Duplicate from akhaliq/Real-Time-Voice-Cloning

Co-authored-by: AK <akhaliq@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,27 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,39 @@
+---
+title: Real Time Voice Cloning
+emoji: 📈
+colorFrom: blue
+colorTo: red
+sdk: gradio
+app_file: app.py
+sdk_version: 3.17.1
+pinned: false
+duplicated_from: akhaliq/Real-Time-Voice-Cloning
+---
+
+# Configuration
+
+`title`: _string_  
+Display title for the Space
+
+`emoji`: _string_  
+Space emoji (emoji-only character allowed)
+
+`colorFrom`: _string_  
+Color for Thumbnail gradient (red, yellow, green, blue, indigo, purple, pink, gray)
+
+`colorTo`: _string_  
+Color for Thumbnail gradient (red, yellow, green, blue, indigo, purple, pink, gray)
+
+`sdk`: _string_  
+Can be either `gradio` or `streamlit`
+
+`sdk_version` : _string_  
+Only applicable for `streamlit` SDK.  
+See [doc](https://hf.co/docs/hub/spaces) for more info on supported versions.
+
+`app_file`: _string_  
+Path to your main application file (which contains either `gradio` or `streamlit` Python code).  
+Path is relative to the root of the repository.
+
+`pinned`: _boolean_  
+Whether the Space stays on top of your list.

app.py

@@ -0,0 +1,61 @@
+import gradio as gr
+import os
+import shlex
+import gdown
+import uuid
+import torch
+
+cpu_param = "--cpu" if not torch.cuda.is_available() else ""
+
+if (not os.path.exists("synpretrained.pt")):
+    gdown.download("https://drive.google.com/u/0/uc?id=1EqFMIbvxffxtjiVrtykroF6_mUh-5Z3s&export=download&confirm=t",
+                   "synpretrained.pt", quiet=False)
+    gdown.download("https://drive.google.com/uc?export=download&id=1q8mEGwCkFy23KZsinbuvdKAQLqNKbYf1",
+                   "encpretrained.pt", quiet=False)
+    gdown.download("https://drive.google.com/uc?export=download&id=1cf2NO6FtI0jDuy8AV3Xgn6leO6dHjIgu",
+                   "vocpretrained.pt", quiet=False)
+
+
+def inference(audio_path, text, mic_path=None):
+    if mic_path:
+        audio_path = mic_path
+    output_path = f"/tmp/output_{uuid.uuid4()}.wav"
+    os.system(
+        f"python demo_cli.py --no_sound {cpu_param} --audio_path {audio_path} --text {shlex.quote(text.strip())} --output_path {output_path}")
+    return output_path
+
+
+title = "Real-Time-Voice-Cloning"
+description = "Gradio demo for Real-Time-Voice-Cloning: Clone a voice in 5 seconds to generate arbitrary speech in real-time. To use it, simply upload your audio, or click one of the examples to load them. Read more at the links below."
+article = "<p style='text-align: center'><a href='https://matheo.uliege.be/handle/2268.2/6801' target='_blank'>Real-Time Voice Cloning</a> | <a href='https://github.com/CorentinJ/Real-Time-Voice-Cloning' target='_blank'>Github Repo</a></p>"
+
+examples = [['test.wav', "This is real time voice cloning on huggingface spaces"]]
+
+
+def toggle(choice):
+    if choice == "mic":
+        return gr.update(visible=True, value=None), gr.update(visible=False, value=None)
+    else:
+        return gr.update(visible=False, value=None), gr.update(visible=True, value=None)
+
+
+with gr.Blocks() as demo:
+    with gr.Row():
+        with gr.Column():
+            radio = gr.Radio(["mic", "file"], value="mic",
+                             label="How would you like to upload your audio?")
+            mic_input = gr.Mic(label="Input", type="filepath", visible=False)
+            audio_file = gr.Audio(
+                type="filepath", label="Input", visible=True)
+            text_input = gr.Textbox(label="Text")
+        with gr.Column():
+            audio_output = gr.Audio(label="Output")
+
+    gr.Examples(examples, fn=inference, inputs=[audio_file, text_input],
+                      outputs=audio_output, cache_examples=True)
+    btn = gr.Button("Generate")
+    btn.click(inference, inputs=[audio_file,
+              text_input, mic_input], outputs=audio_output)
+    radio.change(toggle, radio, [mic_input, audio_file])
+
+demo.launch(enable_queue=True)

demo_cli.py

@@ -0,0 +1,206 @@
+from encoder.params_model import model_embedding_size as speaker_embedding_size
+from utils.argutils import print_args
+from utils.modelutils import check_model_paths
+from synthesizer.inference import Synthesizer
+from encoder import inference as encoder
+from vocoder import inference as vocoder
+from pathlib import Path
+import numpy as np
+import soundfile as sf
+import librosa
+import argparse
+import torch
+import sys
+import os
+from audioread.exceptions import NoBackendError
+
+
+if __name__ == '__main__':
+    ## Info & args
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument("-e", "--enc_model_fpath", type=Path, 
+                        default="encpretrained.pt",
+                        help="Path to a saved encoder")
+    parser.add_argument("-s", "--syn_model_fpath", type=Path, 
+                        default="synpretrained.pt",
+                        help="Path to a saved synthesizer")
+    parser.add_argument("-v", "--voc_model_fpath", type=Path, 
+                        default="vocpretrained.pt",
+                        help="Path to a saved vocoder")
+    parser.add_argument("--cpu", action="store_true", help="If True, processing is done on CPU, even when a GPU is available.")
+    parser.add_argument("--no_sound", action="store_true", help="If True, audio won't be played.")
+    parser.add_argument("--seed", type=int, default=None, help="Optional random number seed value to make toolbox deterministic.")
+    parser.add_argument("--no_mp3_support", action="store_true", help="If True, disallows loading mp3 files to prevent audioread errors when ffmpeg is not installed.")
+    parser.add_argument("-audio", "--audio_path", type=Path, required = True,
+                        help="Path to a audio file")
+    parser.add_argument("--text", type=str, required = True, help="Text Input")
+    parser.add_argument("--output_path", type=str, required = True, help="output file path")
+
+    args = parser.parse_args()
+    print_args(args, parser)
+    if not args.no_sound:
+        import sounddevice as sd
+
+    if args.cpu:
+        # Hide GPUs from Pytorch to force CPU processing
+        os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+
+    if not args.no_mp3_support:
+        try:
+            librosa.load("samples/1320_00000.mp3")
+        except NoBackendError:
+            print("Librosa will be unable to open mp3 files if additional software is not installed.\n"
+                  "Please install ffmpeg or add the '--no_mp3_support' option to proceed without support for mp3 files.")
+            exit(-1)
+        
+    print("Running a test of your configuration...\n")
+        
+    if torch.cuda.is_available():
+        device_id = torch.cuda.current_device()
+        gpu_properties = torch.cuda.get_device_properties(device_id)
+        ## Print some environment information (for debugging purposes)
+        print("Found %d GPUs available. Using GPU %d (%s) of compute capability %d.%d with "
+            "%.1fGb total memory.\n" % 
+            (torch.cuda.device_count(),
+            device_id,
+            gpu_properties.name,
+            gpu_properties.major,
+            gpu_properties.minor,
+            gpu_properties.total_memory / 1e9))
+    else:
+        print("Using CPU for inference.\n")
+    
+    ## Remind the user to download pretrained models if needed
+    check_model_paths(encoder_path=args.enc_model_fpath,
+                      synthesizer_path=args.syn_model_fpath,
+                      vocoder_path=args.voc_model_fpath)
+    
+    ## Load the models one by one.
+    print("Preparing the encoder, the synthesizer and the vocoder...")
+    encoder.load_model(args.enc_model_fpath)
+    synthesizer = Synthesizer(args.syn_model_fpath)
+    vocoder.load_model(args.voc_model_fpath)
+    
+    
+    ## Run a test
+    # print("Testing your configuration with small inputs.")
+    # # Forward an audio waveform of zeroes that lasts 1 second. Notice how we can get the encoder's
+    # # sampling rate, which may differ.
+    # # If you're unfamiliar with digital audio, know that it is encoded as an array of floats 
+    # # (or sometimes integers, but mostly floats in this projects) ranging from -1 to 1.
+    # # The sampling rate is the number of values (samples) recorded per second, it is set to
+    # # 16000 for the encoder. Creating an array of length <sampling_rate> will always correspond 
+    # # to an audio of 1 second.
+    # print("	Testing the encoder...")
+    # encoder.embed_utterance(np.zeros(encoder.sampling_rate))
+    
+    # # Create a dummy embedding. You would normally use the embedding that encoder.embed_utterance
+    # # returns, but here we're going to make one ourselves just for the sake of showing that it's
+    # # possible.
+    # embed = np.random.rand(speaker_embedding_size)
+    # # Embeddings are L2-normalized (this isn't important here, but if you want to make your own 
+    # # embeddings it will be).
+    # embed /= np.linalg.norm(embed)
+    # # The synthesizer can handle multiple inputs with batching. Let's create another embedding to 
+    # # illustrate that
+    # embeds = [embed, np.zeros(speaker_embedding_size)]
+    # texts = ["test 1", "test 2"]
+    # print("	Testing the synthesizer... (loading the model will output a lot of text)")
+    # mels = synthesizer.synthesize_spectrograms(texts, embeds)
+    
+    # # The vocoder synthesizes one waveform at a time, but it's more efficient for long ones. We 
+    # # can concatenate the mel spectrograms to a single one.
+    # mel = np.concatenate(mels, axis=1)
+    # # The vocoder can take a callback function to display the generation. More on that later. For 
+    # # now we'll simply hide it like this:
+    # no_action = lambda *args: None
+    # print("	Testing the vocoder...")
+    # # For the sake of making this test short, we'll pass a short target length. The target length 
+    # # is the length of the wav segments that are processed in parallel. E.g. for audio sampled 
+    # # at 16000 Hertz, a target length of 8000 means that the target audio will be cut in chunks of
+    # # 0.5 seconds which will all be generated together. The parameters here are absurdly short, and 
+    # # that has a detrimental effect on the quality of the audio. The default parameters are 
+    # # recommended in general.
+    # vocoder.infer_waveform(mel, target=200, overlap=50, progress_callback=no_action)
+    
+    print("All test passed! You can now synthesize speech.\n\n")
+    
+    
+    ## Interactive speech generation
+    print("This is a GUI-less example of interface to SV2TTS. The purpose of this script is to "
+          "show how you can interface this project easily with your own. See the source code for "
+          "an explanation of what is happening.\n")
+    
+    print("Interactive generation loop")
+    # while True:
+    # Get the reference audio filepath
+    message = "Reference voice: enter an audio filepath of a voice to be cloned (mp3, "                   "wav, m4a, flac, ...):\n"
+    in_fpath = args.audio_path
+
+    if in_fpath.suffix.lower() == ".mp3" and args.no_mp3_support:
+        print("Can't Use mp3 files please try again:")
+    ## Computing the embedding
+    # First, we load the wav using the function that the speaker encoder provides. This is 
+    # important: there is preprocessing that must be applied.
+    
+    # The following two methods are equivalent:
+    # - Directly load from the filepath:
+    preprocessed_wav = encoder.preprocess_wav(in_fpath)
+    # - If the wav is already loaded:
+    original_wav, sampling_rate = librosa.load(str(in_fpath))
+    preprocessed_wav = encoder.preprocess_wav(original_wav, sampling_rate)
+    print("Loaded file succesfully")
+    
+    # Then we derive the embedding. There are many functions and parameters that the 
+    # speaker encoder interfaces. These are mostly for in-depth research. You will typically
+    # only use this function (with its default parameters):
+    embed = encoder.embed_utterance(preprocessed_wav)
+    print("Created the embedding")
+    
+    
+    ## Generating the spectrogram
+    text = args.text
+    
+    # If seed is specified, reset torch seed and force synthesizer reload
+    if args.seed is not None:
+        torch.manual_seed(args.seed)
+        synthesizer = Synthesizer(args.syn_model_fpath)
+
+    # The synthesizer works in batch, so you need to put your data in a list or numpy array
+    texts = [text]
+    embeds = [embed]
+    # If you know what the attention layer alignments are, you can retrieve them here by
+    # passing return_alignments=True
+    specs = synthesizer.synthesize_spectrograms(texts, embeds)
+    spec = specs[0]
+    print("Created the mel spectrogram")
+    
+    
+    ## Generating the waveform
+    print("Synthesizing the waveform:")
+
+    # If seed is specified, reset torch seed and reload vocoder
+    if args.seed is not None:
+        torch.manual_seed(args.seed)
+        vocoder.load_model(args.voc_model_fpath)
+
+    # Synthesizing the waveform is fairly straightforward. Remember that the longer the
+    # spectrogram, the more time-efficient the vocoder.
+    generated_wav = vocoder.infer_waveform(spec)
+    
+    
+    ## Post-generation
+    # There's a bug with sounddevice that makes the audio cut one second earlier, so we
+    # pad it.
+    generated_wav = np.pad(generated_wav, (0, synthesizer.sample_rate), mode="constant")
+
+    # Trim excess silences to compensate for gaps in spectrograms (issue #53)
+    generated_wav = encoder.preprocess_wav(generated_wav)
+        
+    # Save it on the disk
+    filename = args.output_path
+    print(generated_wav.dtype)
+    sf.write(filename, generated_wav.astype(np.float32), synthesizer.sample_rate)
+    print("\nSaved output as %s\n\n" % filename)

demo_toolbox.py

@@ -0,0 +1,43 @@
+from pathlib import Path
+from toolbox import Toolbox
+from utils.argutils import print_args
+from utils.modelutils import check_model_paths
+import argparse
+import os
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description="Runs the toolbox",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    
+    parser.add_argument("-d", "--datasets_root", type=Path, help= \
+        "Path to the directory containing your datasets. See toolbox/__init__.py for a list of "
+        "supported datasets.", default=None)
+    parser.add_argument("-e", "--enc_models_dir", type=Path, default="encoder/saved_models", 
+                        help="Directory containing saved encoder models")
+    parser.add_argument("-s", "--syn_models_dir", type=Path, default="synthesizer/saved_models", 
+                        help="Directory containing saved synthesizer models")
+    parser.add_argument("-v", "--voc_models_dir", type=Path, default="vocoder/saved_models", 
+                        help="Directory containing saved vocoder models")
+    parser.add_argument("--cpu", action="store_true", help=\
+        "If True, processing is done on CPU, even when a GPU is available.")
+    parser.add_argument("--seed", type=int, default=None, help=\
+        "Optional random number seed value to make toolbox deterministic.")
+    parser.add_argument("--no_mp3_support", action="store_true", help=\
+        "If True, no mp3 files are allowed.")
+    args = parser.parse_args()
+    print_args(args, parser)
+
+    if args.cpu:
+        # Hide GPUs from Pytorch to force CPU processing
+        os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+    del args.cpu
+
+    ## Remind the user to download pretrained models if needed
+    check_model_paths(encoder_path=args.enc_models_dir, synthesizer_path=args.syn_models_dir,
+                      vocoder_path=args.voc_models_dir)
+
+    # Launch the toolbox
+    Toolbox(**vars(args))    

encoder/audio.py

@@ -0,0 +1,117 @@
+from scipy.ndimage.morphology import binary_dilation
+from encoder.params_data import *
+from pathlib import Path
+from typing import Optional, Union
+from warnings import warn
+import numpy as np
+import librosa
+import struct
+
+try:
+    import webrtcvad
+except:
+    warn("Unable to import 'webrtcvad'. This package enables noise removal and is recommended.")
+    webrtcvad=None
+
+int16_max = (2 ** 15) - 1
+
+
+def preprocess_wav(fpath_or_wav: Union[str, Path, np.ndarray],
+                   source_sr: Optional[int] = None,
+                   normalize: Optional[bool] = True,
+                   trim_silence: Optional[bool] = True):
+    """
+    Applies the preprocessing operations used in training the Speaker Encoder to a waveform 
+    either on disk or in memory. The waveform will be resampled to match the data hyperparameters.
+
+    :param fpath_or_wav: either a filepath to an audio file (many extensions are supported, not 
+    just .wav), either the waveform as a numpy array of floats.
+    :param source_sr: if passing an audio waveform, the sampling rate of the waveform before 
+    preprocessing. After preprocessing, the waveform's sampling rate will match the data 
+    hyperparameters. If passing a filepath, the sampling rate will be automatically detected and 
+    this argument will be ignored.
+    """
+    # Load the wav from disk if needed
+    if isinstance(fpath_or_wav, str) or isinstance(fpath_or_wav, Path):
+        wav, source_sr = librosa.load(str(fpath_or_wav), sr=None)
+    else:
+        wav = fpath_or_wav
+    
+    # Resample the wav if needed
+    if source_sr is not None and source_sr != sampling_rate:
+        wav = librosa.resample(wav, source_sr, sampling_rate)
+        
+    # Apply the preprocessing: normalize volume and shorten long silences 
+    if normalize:
+        wav = normalize_volume(wav, audio_norm_target_dBFS, increase_only=True)
+    if webrtcvad and trim_silence:
+        wav = trim_long_silences(wav)
+    
+    return wav
+
+
+def wav_to_mel_spectrogram(wav):
+    """
+    Derives a mel spectrogram ready to be used by the encoder from a preprocessed audio waveform.
+    Note: this not a log-mel spectrogram.
+    """
+    frames = librosa.feature.melspectrogram(
+        wav,
+        sampling_rate,
+        n_fft=int(sampling_rate * mel_window_length / 1000),
+        hop_length=int(sampling_rate * mel_window_step / 1000),
+        n_mels=mel_n_channels
+    )
+    return frames.astype(np.float32).T
+
+
+def trim_long_silences(wav):
+    """
+    Ensures that segments without voice in the waveform remain no longer than a 
+    threshold determined by the VAD parameters in params.py.
+
+    :param wav: the raw waveform as a numpy array of floats 
+    :return: the same waveform with silences trimmed away (length <= original wav length)
+    """
+    # Compute the voice detection window size
+    samples_per_window = (vad_window_length * sampling_rate) // 1000
+    
+    # Trim the end of the audio to have a multiple of the window size
+    wav = wav[:len(wav) - (len(wav) % samples_per_window)]
+    
+    # Convert the float waveform to 16-bit mono PCM
+    pcm_wave = struct.pack("%dh" % len(wav), *(np.round(wav * int16_max)).astype(np.int16))
+    
+    # Perform voice activation detection
+    voice_flags = []
+    vad = webrtcvad.Vad(mode=3)
+    for window_start in range(0, len(wav), samples_per_window):
+        window_end = window_start + samples_per_window
+        voice_flags.append(vad.is_speech(pcm_wave[window_start * 2:window_end * 2],
+                                         sample_rate=sampling_rate))
+    voice_flags = np.array(voice_flags)
+    
+    # Smooth the voice detection with a moving average
+    def moving_average(array, width):
+        array_padded = np.concatenate((np.zeros((width - 1) // 2), array, np.zeros(width // 2)))
+        ret = np.cumsum(array_padded, dtype=float)
+        ret[width:] = ret[width:] - ret[:-width]
+        return ret[width - 1:] / width
+    
+    audio_mask = moving_average(voice_flags, vad_moving_average_width)
+    audio_mask = np.round(audio_mask).astype(np.bool)
+    
+    # Dilate the voiced regions
+    audio_mask = binary_dilation(audio_mask, np.ones(vad_max_silence_length + 1))
+    audio_mask = np.repeat(audio_mask, samples_per_window)
+    
+    return wav[audio_mask == True]
+
+
+def normalize_volume(wav, target_dBFS, increase_only=False, decrease_only=False):
+    if increase_only and decrease_only:
+        raise ValueError("Both increase only and decrease only are set")
+    dBFS_change = target_dBFS - 10 * np.log10(np.mean(wav ** 2))
+    if (dBFS_change < 0 and increase_only) or (dBFS_change > 0 and decrease_only):
+        return wav
+    return wav * (10 ** (dBFS_change / 20))

encoder/config.py

@@ -0,0 +1,45 @@
+librispeech_datasets = {
+    "train": {
+        "clean": ["LibriSpeech/train-clean-100", "LibriSpeech/train-clean-360"],
+        "other": ["LibriSpeech/train-other-500"]
+    },
+    "test": {
+        "clean": ["LibriSpeech/test-clean"],
+        "other": ["LibriSpeech/test-other"]
+    },
+    "dev": {
+        "clean": ["LibriSpeech/dev-clean"],
+        "other": ["LibriSpeech/dev-other"]
+    },
+}
+libritts_datasets = {
+    "train": {
+        "clean": ["LibriTTS/train-clean-100", "LibriTTS/train-clean-360"],
+        "other": ["LibriTTS/train-other-500"]
+    },
+    "test": {
+        "clean": ["LibriTTS/test-clean"],
+        "other": ["LibriTTS/test-other"]
+    },
+    "dev": {
+        "clean": ["LibriTTS/dev-clean"],
+        "other": ["LibriTTS/dev-other"]
+    },
+}
+voxceleb_datasets = {
+    "voxceleb1" : {
+        "train": ["VoxCeleb1/wav"],
+        "test": ["VoxCeleb1/test_wav"]
+    },
+    "voxceleb2" : {
+        "train": ["VoxCeleb2/dev/aac"],
+        "test": ["VoxCeleb2/test_wav"]
+    }
+}
+
+other_datasets = [
+    "LJSpeech-1.1",
+    "VCTK-Corpus/wav48",
+]
+
+anglophone_nationalites = ["australia", "canada", "ireland", "uk", "usa"]

encoder/data_objects/__init__.py

@@ -0,0 +1,2 @@
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataset
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataLoader

encoder/data_objects/random_cycler.py

@@ -0,0 +1,37 @@
+import random
+
+class RandomCycler:
+    """
+    Creates an internal copy of a sequence and allows access to its items in a constrained random 
+    order. For a source sequence of n items and one or several consecutive queries of a total 
+    of m items, the following guarantees hold (one implies the other):
+        - Each item will be returned between m // n and ((m - 1) // n) + 1 times.
+        - Between two appearances of the same item, there may be at most 2 * (n - 1) other items.
+    """
+    
+    def __init__(self, source):
+        if len(source) == 0:
+            raise Exception("Can't create RandomCycler from an empty collection")
+        self.all_items = list(source)
+        self.next_items = []
+    
+    def sample(self, count: int):
+        shuffle = lambda l: random.sample(l, len(l))
+        
+        out = []
+        while count > 0:
+            if count >= len(self.all_items):
+                out.extend(shuffle(list(self.all_items)))
+                count -= len(self.all_items)
+                continue
+            n = min(count, len(self.next_items))
+            out.extend(self.next_items[:n])
+            count -= n
+            self.next_items = self.next_items[n:]
+            if len(self.next_items) == 0:
+                self.next_items = shuffle(list(self.all_items))
+        return out
+    
+    def __next__(self):
+        return self.sample(1)[0]
+

encoder/data_objects/speaker.py

@@ -0,0 +1,40 @@
+from encoder.data_objects.random_cycler import RandomCycler
+from encoder.data_objects.utterance import Utterance
+from pathlib import Path
+
+# Contains the set of utterances of a single speaker
+class Speaker:
+    def __init__(self, root: Path):
+        self.root = root
+        self.name = root.name
+        self.utterances = None
+        self.utterance_cycler = None
+        
+    def _load_utterances(self):
+        with self.root.joinpath("_sources.txt").open("r") as sources_file:
+            sources = [l.split(",") for l in sources_file]
+        sources = {frames_fname: wave_fpath for frames_fname, wave_fpath in sources}
+        self.utterances = [Utterance(self.root.joinpath(f), w) for f, w in sources.items()]
+        self.utterance_cycler = RandomCycler(self.utterances)
+               
+    def random_partial(self, count, n_frames):
+        """
+        Samples a batch of <count> unique partial utterances from the disk in a way that all 
+        utterances come up at least once every two cycles and in a random order every time.
+        
+        :param count: The number of partial utterances to sample from the set of utterances from 
+        that speaker. Utterances are guaranteed not to be repeated if <count> is not larger than 
+        the number of utterances available.
+        :param n_frames: The number of frames in the partial utterance.
+        :return: A list of tuples (utterance, frames, range) where utterance is an Utterance, 
+        frames are the frames of the partial utterances and range is the range of the partial 
+        utterance with regard to the complete utterance.
+        """
+        if self.utterances is None:
+            self._load_utterances()
+
+        utterances = self.utterance_cycler.sample(count)
+
+        a = [(u,) + u.random_partial(n_frames) for u in utterances]
+
+        return a

encoder/data_objects/speaker_batch.py

@@ -0,0 +1,12 @@
+import numpy as np
+from typing import List
+from encoder.data_objects.speaker import Speaker
+
+class SpeakerBatch:
+    def __init__(self, speakers: List[Speaker], utterances_per_speaker: int, n_frames: int):
+        self.speakers = speakers
+        self.partials = {s: s.random_partial(utterances_per_speaker, n_frames) for s in speakers}
+        
+        # Array of shape (n_speakers * n_utterances, n_frames, mel_n), e.g. for 3 speakers with
+        # 4 utterances each of 160 frames of 40 mel coefficients: (12, 160, 40)
+        self.data = np.array([frames for s in speakers for _, frames, _ in self.partials[s]])

encoder/data_objects/speaker_verification_dataset.py

@@ -0,0 +1,56 @@
+from encoder.data_objects.random_cycler import RandomCycler
+from encoder.data_objects.speaker_batch import SpeakerBatch
+from encoder.data_objects.speaker import Speaker
+from encoder.params_data import partials_n_frames
+from torch.utils.data import Dataset, DataLoader
+from pathlib import Path
+
+# TODO: improve with a pool of speakers for data efficiency
+
+class SpeakerVerificationDataset(Dataset):
+    def __init__(self, datasets_root: Path):
+        self.root = datasets_root
+        speaker_dirs = [f for f in self.root.glob("*") if f.is_dir()]
+        if len(speaker_dirs) == 0:
+            raise Exception("No speakers found. Make sure you are pointing to the directory "
+                            "containing all preprocessed speaker directories.")
+        self.speakers = [Speaker(speaker_dir) for speaker_dir in speaker_dirs]
+        self.speaker_cycler = RandomCycler(self.speakers)
+
+    def __len__(self):
+        return int(1e10)
+        
+    def __getitem__(self, index):
+        return next(self.speaker_cycler)
+    
+    def get_logs(self):
+        log_string = ""
+        for log_fpath in self.root.glob("*.txt"):
+            with log_fpath.open("r") as log_file:
+                log_string += "".join(log_file.readlines())
+        return log_string
+    
+    
+class SpeakerVerificationDataLoader(DataLoader):
+    def __init__(self, dataset, speakers_per_batch, utterances_per_speaker, sampler=None, 
+                 batch_sampler=None, num_workers=0, pin_memory=False, timeout=0, 
+                 worker_init_fn=None):
+        self.utterances_per_speaker = utterances_per_speaker
+
+        super().__init__(
+            dataset=dataset, 
+            batch_size=speakers_per_batch, 
+            shuffle=False, 
+            sampler=sampler, 
+            batch_sampler=batch_sampler, 
+            num_workers=num_workers,
+            collate_fn=self.collate, 
+            pin_memory=pin_memory, 
+            drop_last=False, 
+            timeout=timeout, 
+            worker_init_fn=worker_init_fn
+        )
+
+    def collate(self, speakers):
+        return SpeakerBatch(speakers, self.utterances_per_speaker, partials_n_frames) 
+    

encoder/data_objects/utterance.py

@@ -0,0 +1,26 @@
+import numpy as np
+
+
+class Utterance:
+    def __init__(self, frames_fpath, wave_fpath):
+        self.frames_fpath = frames_fpath
+        self.wave_fpath = wave_fpath
+        
+    def get_frames(self):
+        return np.load(self.frames_fpath)
+
+    def random_partial(self, n_frames):
+        """
+        Crops the frames into a partial utterance of n_frames
+        
+        :param n_frames: The number of frames of the partial utterance
+        :return: the partial utterance frames and a tuple indicating the start and end of the 
+        partial utterance in the complete utterance.
+        """
+        frames = self.get_frames()
+        if frames.shape[0] == n_frames:
+            start = 0
+        else:
+            start = np.random.randint(0, frames.shape[0] - n_frames)
+        end = start + n_frames
+        return frames[start:end], (start, end)

encoder/inference.py

@@ -0,0 +1,178 @@
+from encoder.params_data import *
+from encoder.model import SpeakerEncoder
+from encoder.audio import preprocess_wav   # We want to expose this function from here
+from matplotlib import cm
+from encoder import audio
+from pathlib import Path
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+_model = None # type: SpeakerEncoder
+_device = None # type: torch.device
+
+
+def load_model(weights_fpath: Path, device=None):
+    """
+    Loads the model in memory. If this function is not explicitely called, it will be run on the 
+    first call to embed_frames() with the default weights file.
+    
+    :param weights_fpath: the path to saved model weights.
+    :param device: either a torch device or the name of a torch device (e.g. "cpu", "cuda"). The 
+    model will be loaded and will run on this device. Outputs will however always be on the cpu. 
+    If None, will default to your GPU if it"s available, otherwise your CPU.
+    """
+    # TODO: I think the slow loading of the encoder might have something to do with the device it
+    #   was saved on. Worth investigating.
+    global _model, _device
+    if device is None:
+        _device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    elif isinstance(device, str):
+        _device = torch.device(device)
+    _model = SpeakerEncoder(_device, torch.device("cpu"))
+    checkpoint = torch.load(weights_fpath, _device)
+    _model.load_state_dict(checkpoint["model_state"])
+    _model.eval()
+    print("Loaded encoder \"%s\" trained to step %d" % (weights_fpath.name, checkpoint["step"]))
+    
+    
+def is_loaded():
+    return _model is not None
+
+
+def embed_frames_batch(frames_batch):
+    """
+    Computes embeddings for a batch of mel spectrogram.
+    
+    :param frames_batch: a batch mel of spectrogram as a numpy array of float32 of shape 
+    (batch_size, n_frames, n_channels)
+    :return: the embeddings as a numpy array of float32 of shape (batch_size, model_embedding_size)
+    """
+    if _model is None:
+        raise Exception("Model was not loaded. Call load_model() before inference.")
+    
+    frames = torch.from_numpy(frames_batch).to(_device)
+    embed = _model.forward(frames).detach().cpu().numpy()
+    return embed
+
+
+def compute_partial_slices(n_samples, partial_utterance_n_frames=partials_n_frames,
+                           min_pad_coverage=0.75, overlap=0.5):
+    """
+    Computes where to split an utterance waveform and its corresponding mel spectrogram to obtain 
+    partial utterances of <partial_utterance_n_frames> each. Both the waveform and the mel 
+    spectrogram slices are returned, so as to make each partial utterance waveform correspond to 
+    its spectrogram. This function assumes that the mel spectrogram parameters used are those 
+    defined in params_data.py.
+    
+    The returned ranges may be indexing further than the length of the waveform. It is 
+    recommended that you pad the waveform with zeros up to wave_slices[-1].stop.
+    
+    :param n_samples: the number of samples in the waveform
+    :param partial_utterance_n_frames: the number of mel spectrogram frames in each partial 
+    utterance
+    :param min_pad_coverage: when reaching the last partial utterance, it may or may not have 
+    enough frames. If at least <min_pad_coverage> of <partial_utterance_n_frames> are present, 
+    then the last partial utterance will be considered, as if we padded the audio. Otherwise, 
+    it will be discarded, as if we trimmed the audio. If there aren't enough frames for 1 partial 
+    utterance, this parameter is ignored so that the function always returns at least 1 slice.
+    :param overlap: by how much the partial utterance should overlap. If set to 0, the partial 
+    utterances are entirely disjoint. 
+    :return: the waveform slices and mel spectrogram slices as lists of array slices. Index 
+    respectively the waveform and the mel spectrogram with these slices to obtain the partial 
+    utterances.
+    """
+    assert 0 <= overlap < 1
+    assert 0 < min_pad_coverage <= 1
+    
+    samples_per_frame = int((sampling_rate * mel_window_step / 1000))
+    n_frames = int(np.ceil((n_samples + 1) / samples_per_frame))
+    frame_step = max(int(np.round(partial_utterance_n_frames * (1 - overlap))), 1)
+
+    # Compute the slices
+    wav_slices, mel_slices = [], []
+    steps = max(1, n_frames - partial_utterance_n_frames + frame_step + 1)
+    for i in range(0, steps, frame_step):
+        mel_range = np.array([i, i + partial_utterance_n_frames])
+        wav_range = mel_range * samples_per_frame
+        mel_slices.append(slice(*mel_range))
+        wav_slices.append(slice(*wav_range))
+        
+    # Evaluate whether extra padding is warranted or not
+    last_wav_range = wav_slices[-1]
+    coverage = (n_samples - last_wav_range.start) / (last_wav_range.stop - last_wav_range.start)
+    if coverage < min_pad_coverage and len(mel_slices) > 1:
+        mel_slices = mel_slices[:-1]
+        wav_slices = wav_slices[:-1]
+    
+    return wav_slices, mel_slices
+
+
+def embed_utterance(wav, using_partials=True, return_partials=False, **kwargs):
+    """
+    Computes an embedding for a single utterance.
+    
+    # TODO: handle multiple wavs to benefit from batching on GPU
+    :param wav: a preprocessed (see audio.py) utterance waveform as a numpy array of float32
+    :param using_partials: if True, then the utterance is split in partial utterances of 
+    <partial_utterance_n_frames> frames and the utterance embedding is computed from their 
+    normalized average. If False, the utterance is instead computed from feeding the entire 
+    spectogram to the network.
+    :param return_partials: if True, the partial embeddings will also be returned along with the 
+    wav slices that correspond to the partial embeddings.
+    :param kwargs: additional arguments to compute_partial_splits()
+    :return: the embedding as a numpy array of float32 of shape (model_embedding_size,). If 
+    <return_partials> is True, the partial utterances as a numpy array of float32 of shape 
+    (n_partials, model_embedding_size) and the wav partials as a list of slices will also be 
+    returned. If <using_partials> is simultaneously set to False, both these values will be None 
+    instead.
+    """
+    # Process the entire utterance if not using partials
+    if not using_partials:
+        frames = audio.wav_to_mel_spectrogram(wav)
+        embed = embed_frames_batch(frames[None, ...])[0]
+        if return_partials:
+            return embed, None, None
+        return embed
+    
+    # Compute where to split the utterance into partials and pad if necessary
+    wave_slices, mel_slices = compute_partial_slices(len(wav), **kwargs)
+    max_wave_length = wave_slices[-1].stop
+    if max_wave_length >= len(wav):
+        wav = np.pad(wav, (0, max_wave_length - len(wav)), "constant")
+    
+    # Split the utterance into partials
+    frames = audio.wav_to_mel_spectrogram(wav)
+    frames_batch = np.array([frames[s] for s in mel_slices])
+    partial_embeds = embed_frames_batch(frames_batch)
+    
+    # Compute the utterance embedding from the partial embeddings
+    raw_embed = np.mean(partial_embeds, axis=0)
+    embed = raw_embed / np.linalg.norm(raw_embed, 2)
+    
+    if return_partials:
+        return embed, partial_embeds, wave_slices
+    return embed
+
+
+def embed_speaker(wavs, **kwargs):
+    raise NotImplemented()
+
+
+def plot_embedding_as_heatmap(embed, ax=None, title="", shape=None, color_range=(0, 0.30)):
+    if ax is None:
+        ax = plt.gca()
+    
+    if shape is None:
+        height = int(np.sqrt(len(embed)))
+        shape = (height, -1)
+    embed = embed.reshape(shape)
+    
+    cmap = cm.get_cmap()
+    mappable = ax.imshow(embed, cmap=cmap)
+    cbar = plt.colorbar(mappable, ax=ax, fraction=0.046, pad=0.04)
+    sm = cm.ScalarMappable(cmap=cmap)
+    sm.set_clim(*color_range)
+    
+    ax.set_xticks([]), ax.set_yticks([])
+    ax.set_title(title)

encoder/model.py

@@ -0,0 +1,135 @@
+from encoder.params_model import *
+from encoder.params_data import *
+from scipy.interpolate import interp1d
+from sklearn.metrics import roc_curve
+from torch.nn.utils import clip_grad_norm_
+from scipy.optimize import brentq
+from torch import nn
+import numpy as np
+import torch
+
+
+class SpeakerEncoder(nn.Module):
+    def __init__(self, device, loss_device):
+        super().__init__()
+        self.loss_device = loss_device
+        
+        # Network defition
+        self.lstm = nn.LSTM(input_size=mel_n_channels,
+                            hidden_size=model_hidden_size, 
+                            num_layers=model_num_layers, 
+                            batch_first=True).to(device)
+        self.linear = nn.Linear(in_features=model_hidden_size, 
+                                out_features=model_embedding_size).to(device)
+        self.relu = torch.nn.ReLU().to(device)
+        
+        # Cosine similarity scaling (with fixed initial parameter values)
+        self.similarity_weight = nn.Parameter(torch.tensor([10.])).to(loss_device)
+        self.similarity_bias = nn.Parameter(torch.tensor([-5.])).to(loss_device)
+
+        # Loss
+        self.loss_fn = nn.CrossEntropyLoss().to(loss_device)
+        
+    def do_gradient_ops(self):
+        # Gradient scale
+        self.similarity_weight.grad *= 0.01
+        self.similarity_bias.grad *= 0.01
+            
+        # Gradient clipping
+        clip_grad_norm_(self.parameters(), 3, norm_type=2)
+    
+    def forward(self, utterances, hidden_init=None):
+        """
+        Computes the embeddings of a batch of utterance spectrograms.
+        
+        :param utterances: batch of mel-scale filterbanks of same duration as a tensor of shape 
+        (batch_size, n_frames, n_channels) 
+        :param hidden_init: initial hidden state of the LSTM as a tensor of shape (num_layers, 
+        batch_size, hidden_size). Will default to a tensor of zeros if None.
+        :return: the embeddings as a tensor of shape (batch_size, embedding_size)
+        """
+        # Pass the input through the LSTM layers and retrieve all outputs, the final hidden state
+        # and the final cell state.
+        out, (hidden, cell) = self.lstm(utterances, hidden_init)
+        
+        # We take only the hidden state of the last layer
+        embeds_raw = self.relu(self.linear(hidden[-1]))
+        
+        # L2-normalize it
+        embeds = embeds_raw / (torch.norm(embeds_raw, dim=1, keepdim=True) + 1e-5)        
+
+        return embeds
+    
+    def similarity_matrix(self, embeds):
+        """
+        Computes the similarity matrix according the section 2.1 of GE2E.
+
+        :param embeds: the embeddings as a tensor of shape (speakers_per_batch, 
+        utterances_per_speaker, embedding_size)
+        :return: the similarity matrix as a tensor of shape (speakers_per_batch,
+        utterances_per_speaker, speakers_per_batch)
+        """
+        speakers_per_batch, utterances_per_speaker = embeds.shape[:2]
+        
+        # Inclusive centroids (1 per speaker). Cloning is needed for reverse differentiation
+        centroids_incl = torch.mean(embeds, dim=1, keepdim=True)
+        centroids_incl = centroids_incl.clone() / (torch.norm(centroids_incl, dim=2, keepdim=True) + 1e-5)
+
+        # Exclusive centroids (1 per utterance)
+        centroids_excl = (torch.sum(embeds, dim=1, keepdim=True) - embeds)
+        centroids_excl /= (utterances_per_speaker - 1)
+        centroids_excl = centroids_excl.clone() / (torch.norm(centroids_excl, dim=2, keepdim=True) + 1e-5)
+
+        # Similarity matrix. The cosine similarity of already 2-normed vectors is simply the dot
+        # product of these vectors (which is just an element-wise multiplication reduced by a sum).
+        # We vectorize the computation for efficiency.
+        sim_matrix = torch.zeros(speakers_per_batch, utterances_per_speaker,
+                                 speakers_per_batch).to(self.loss_device)
+        mask_matrix = 1 - np.eye(speakers_per_batch, dtype=np.int)
+        for j in range(speakers_per_batch):
+            mask = np.where(mask_matrix[j])[0]
+            sim_matrix[mask, :, j] = (embeds[mask] * centroids_incl[j]).sum(dim=2)
+            sim_matrix[j, :, j] = (embeds[j] * centroids_excl[j]).sum(dim=1)
+        
+        ## Even more vectorized version (slower maybe because of transpose)
+        # sim_matrix2 = torch.zeros(speakers_per_batch, speakers_per_batch, utterances_per_speaker
+        #                           ).to(self.loss_device)
+        # eye = np.eye(speakers_per_batch, dtype=np.int)
+        # mask = np.where(1 - eye)
+        # sim_matrix2[mask] = (embeds[mask[0]] * centroids_incl[mask[1]]).sum(dim=2)
+        # mask = np.where(eye)
+        # sim_matrix2[mask] = (embeds * centroids_excl).sum(dim=2)
+        # sim_matrix2 = sim_matrix2.transpose(1, 2)
+        
+        sim_matrix = sim_matrix * self.similarity_weight + self.similarity_bias
+        return sim_matrix
+    
+    def loss(self, embeds):
+        """
+        Computes the softmax loss according the section 2.1 of GE2E.
+        
+        :param embeds: the embeddings as a tensor of shape (speakers_per_batch, 
+        utterances_per_speaker, embedding_size)
+        :return: the loss and the EER for this batch of embeddings.
+        """
+        speakers_per_batch, utterances_per_speaker = embeds.shape[:2]
+        
+        # Loss
+        sim_matrix = self.similarity_matrix(embeds)
+        sim_matrix = sim_matrix.reshape((speakers_per_batch * utterances_per_speaker, 
+                                         speakers_per_batch))
+        ground_truth = np.repeat(np.arange(speakers_per_batch), utterances_per_speaker)
+        target = torch.from_numpy(ground_truth).long().to(self.loss_device)
+        loss = self.loss_fn(sim_matrix, target)
+        
+        # EER (not backpropagated)
+        with torch.no_grad():
+            inv_argmax = lambda i: np.eye(1, speakers_per_batch, i, dtype=np.int)[0]
+            labels = np.array([inv_argmax(i) for i in ground_truth])
+            preds = sim_matrix.detach().cpu().numpy()
+
+            # Snippet from https://yangcha.github.io/EER-ROC/
+            fpr, tpr, thresholds = roc_curve(labels.flatten(), preds.flatten())           
+            eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
+            
+        return loss, eer

encoder/params_data.py

@@ -0,0 +1,29 @@
+
+## Mel-filterbank
+mel_window_length = 25  # In milliseconds
+mel_window_step = 10    # In milliseconds
+mel_n_channels = 40
+
+
+## Audio
+sampling_rate = 16000
+# Number of spectrogram frames in a partial utterance
+partials_n_frames = 160     # 1600 ms
+# Number of spectrogram frames at inference
+inference_n_frames = 80     #  800 ms
+
+
+## Voice Activation Detection
+# Window size of the VAD. Must be either 10, 20 or 30 milliseconds.
+# This sets the granularity of the VAD. Should not need to be changed.
+vad_window_length = 30  # In milliseconds
+# Number of frames to average together when performing the moving average smoothing.
+# The larger this value, the larger the VAD variations must be to not get smoothed out. 
+vad_moving_average_width = 8
+# Maximum number of consecutive silent frames a segment can have.
+vad_max_silence_length = 6
+
+
+## Audio volume normalization
+audio_norm_target_dBFS = -30
+

encoder/params_model.py

@@ -0,0 +1,11 @@
+
+## Model parameters
+model_hidden_size = 256
+model_embedding_size = 256
+model_num_layers = 3
+
+
+## Training parameters
+learning_rate_init = 1e-4
+speakers_per_batch = 64
+utterances_per_speaker = 10

encoder/preprocess.py

@@ -0,0 +1,175 @@
+from multiprocess.pool import ThreadPool
+from encoder.params_data import *
+from encoder.config import librispeech_datasets, anglophone_nationalites
+from datetime import datetime
+from encoder import audio
+from pathlib import Path
+from tqdm import tqdm
+import numpy as np
+
+
+class DatasetLog:
+    """
+    Registers metadata about the dataset in a text file.
+    """
+    def __init__(self, root, name):
+        self.text_file = open(Path(root, "Log_%s.txt" % name.replace("/", "_")), "w")
+        self.sample_data = dict()
+        
+        start_time = str(datetime.now().strftime("%A %d %B %Y at %H:%M"))
+        self.write_line("Creating dataset %s on %s" % (name, start_time))
+        self.write_line("-----")
+        self._log_params()
+        
+    def _log_params(self):
+        from encoder import params_data
+        self.write_line("Parameter values:")
+        for param_name in (p for p in dir(params_data) if not p.startswith("__")):
+            value = getattr(params_data, param_name)
+            self.write_line("\t%s: %s" % (param_name, value))
+        self.write_line("-----")
+    
+    def write_line(self, line):
+        self.text_file.write("%s\n" % line)
+        
+    def add_sample(self, **kwargs):
+        for param_name, value in kwargs.items():
+            if not param_name in self.sample_data:
+                self.sample_data[param_name] = []
+            self.sample_data[param_name].append(value)
+            
+    def finalize(self):
+        self.write_line("Statistics:")
+        for param_name, values in self.sample_data.items():
+            self.write_line("\t%s:" % param_name)
+            self.write_line("\t\tmin %.3f, max %.3f" % (np.min(values), np.max(values)))
+            self.write_line("\t\tmean %.3f, median %.3f" % (np.mean(values), np.median(values)))
+        self.write_line("-----")
+        end_time = str(datetime.now().strftime("%A %d %B %Y at %H:%M"))
+        self.write_line("Finished on %s" % end_time)
+        self.text_file.close()
+       
+        
+def _init_preprocess_dataset(dataset_name, datasets_root, out_dir) -> (Path, DatasetLog):
+    dataset_root = datasets_root.joinpath(dataset_name)
+    if not dataset_root.exists():
+        print("Couldn\'t find %s, skipping this dataset." % dataset_root)
+        return None, None
+    return dataset_root, DatasetLog(out_dir, dataset_name)
+
+
+def _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, extension,
+                             skip_existing, logger):
+    print("%s: Preprocessing data for %d speakers." % (dataset_name, len(speaker_dirs)))
+    
+    # Function to preprocess utterances for one speaker
+    def preprocess_speaker(speaker_dir: Path):
+        # Give a name to the speaker that includes its dataset
+        speaker_name = "_".join(speaker_dir.relative_to(datasets_root).parts)
+        
+        # Create an output directory with that name, as well as a txt file containing a 
+        # reference to each source file.
+        speaker_out_dir = out_dir.joinpath(speaker_name)
+        speaker_out_dir.mkdir(exist_ok=True)
+        sources_fpath = speaker_out_dir.joinpath("_sources.txt")
+        
+        # There's a possibility that the preprocessing was interrupted earlier, check if 
+        # there already is a sources file.
+        if sources_fpath.exists():
+            try:
+                with sources_fpath.open("r") as sources_file:
+                    existing_fnames = {line.split(",")[0] for line in sources_file}
+            except:
+                existing_fnames = {}
+        else:
+            existing_fnames = {}
+        
+        # Gather all audio files for that speaker recursively
+        sources_file = sources_fpath.open("a" if skip_existing else "w")
+        for in_fpath in speaker_dir.glob("**/*.%s" % extension):
+            # Check if the target output file already exists
+            out_fname = "_".join(in_fpath.relative_to(speaker_dir).parts)
+            out_fname = out_fname.replace(".%s" % extension, ".npy")
+            if skip_existing and out_fname in existing_fnames:
+                continue
+                
+            # Load and preprocess the waveform
+            wav = audio.preprocess_wav(in_fpath)
+            if len(wav) == 0:
+                continue
+            
+            # Create the mel spectrogram, discard those that are too short
+            frames = audio.wav_to_mel_spectrogram(wav)
+            if len(frames) < partials_n_frames:
+                continue
+            
+            out_fpath = speaker_out_dir.joinpath(out_fname)
+            np.save(out_fpath, frames)
+            logger.add_sample(duration=len(wav) / sampling_rate)
+            sources_file.write("%s,%s\n" % (out_fname, in_fpath))
+        
+        sources_file.close()
+    
+    # Process the utterances for each speaker
+    with ThreadPool(8) as pool:
+        list(tqdm(pool.imap(preprocess_speaker, speaker_dirs), dataset_name, len(speaker_dirs),
+                  unit="speakers"))
+    logger.finalize()
+    print("Done preprocessing %s.\n" % dataset_name)
+
+
+def preprocess_librispeech(datasets_root: Path, out_dir: Path, skip_existing=False):
+    for dataset_name in librispeech_datasets["train"]["other"]:
+        # Initialize the preprocessing
+        dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+        if not dataset_root:
+            return 
+        
+        # Preprocess all speakers
+        speaker_dirs = list(dataset_root.glob("*"))
+        _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, "flac",
+                                 skip_existing, logger)
+
+
+def preprocess_voxceleb1(datasets_root: Path, out_dir: Path, skip_existing=False):
+    # Initialize the preprocessing
+    dataset_name = "VoxCeleb1"
+    dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+    if not dataset_root:
+        return
+
+    # Get the contents of the meta file
+    with dataset_root.joinpath("vox1_meta.csv").open("r") as metafile:
+        metadata = [line.split("\t") for line in metafile][1:]
+    
+    # Select the ID and the nationality, filter out non-anglophone speakers
+    nationalities = {line[0]: line[3] for line in metadata}
+    keep_speaker_ids = [speaker_id for speaker_id, nationality in nationalities.items() if 
+                        nationality.lower() in anglophone_nationalites]
+    print("VoxCeleb1: using samples from %d (presumed anglophone) speakers out of %d." % 
+          (len(keep_speaker_ids), len(nationalities)))
+    
+    # Get the speaker directories for anglophone speakers only
+    speaker_dirs = dataset_root.joinpath("wav").glob("*")
+    speaker_dirs = [speaker_dir for speaker_dir in speaker_dirs if
+                    speaker_dir.name in keep_speaker_ids]
+    print("VoxCeleb1: found %d anglophone speakers on the disk, %d missing (this is normal)." % 
+          (len(speaker_dirs), len(keep_speaker_ids) - len(speaker_dirs)))
+
+    # Preprocess all speakers
+    _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, "wav",
+                             skip_existing, logger)
+
+
+def preprocess_voxceleb2(datasets_root: Path, out_dir: Path, skip_existing=False):
+    # Initialize the preprocessing
+    dataset_name = "VoxCeleb2"
+    dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+    if not dataset_root:
+        return
+    
+    # Get the speaker directories
+    # Preprocess all speakers
+    speaker_dirs = list(dataset_root.joinpath("dev", "aac").glob("*"))
+    _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, "m4a",
+                             skip_existing, logger)

encoder/train.py

@@ -0,0 +1,123 @@
+from encoder.visualizations import Visualizations
+from encoder.data_objects import SpeakerVerificationDataLoader, SpeakerVerificationDataset
+from encoder.params_model import *
+from encoder.model import SpeakerEncoder
+from utils.profiler import Profiler
+from pathlib import Path
+import torch
+
+def sync(device: torch.device):
+    # For correct profiling (cuda operations are async)
+    if device.type == "cuda":
+        torch.cuda.synchronize(device)
+    
+
+def train(run_id: str, clean_data_root: Path, models_dir: Path, umap_every: int, save_every: int,
+          backup_every: int, vis_every: int, force_restart: bool, visdom_server: str,
+          no_visdom: bool):
+    # Create a dataset and a dataloader
+    dataset = SpeakerVerificationDataset(clean_data_root)
+    loader = SpeakerVerificationDataLoader(
+        dataset,
+        speakers_per_batch,
+        utterances_per_speaker,
+        num_workers=8,
+    )
+    
+    # Setup the device on which to run the forward pass and the loss. These can be different, 
+    # because the forward pass is faster on the GPU whereas the loss is often (depending on your
+    # hyperparameters) faster on the CPU.
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    # FIXME: currently, the gradient is None if loss_device is cuda
+    loss_device = torch.device("cpu")
+    
+    # Create the model and the optimizer
+    model = SpeakerEncoder(device, loss_device)
+    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate_init)
+    init_step = 1
+    
+    # Configure file path for the model
+    state_fpath = models_dir.joinpath(run_id + ".pt")
+    backup_dir = models_dir.joinpath(run_id + "_backups")
+
+    # Load any existing model
+    if not force_restart:
+        if state_fpath.exists():
+            print("Found existing model \"%s\", loading it and resuming training." % run_id)
+            checkpoint = torch.load(state_fpath)
+            init_step = checkpoint["step"]
+            model.load_state_dict(checkpoint["model_state"])
+            optimizer.load_state_dict(checkpoint["optimizer_state"])
+            optimizer.param_groups[0]["lr"] = learning_rate_init
+        else:
+            print("No model \"%s\" found, starting training from scratch." % run_id)
+    else:
+        print("Starting the training from scratch.")
+    model.train()
+    
+    # Initialize the visualization environment
+    vis = Visualizations(run_id, vis_every, server=visdom_server, disabled=no_visdom)
+    vis.log_dataset(dataset)
+    vis.log_params()
+    device_name = str(torch.cuda.get_device_name(0) if torch.cuda.is_available() else "CPU")
+    vis.log_implementation({"Device": device_name})
+    
+    # Training loop
+    profiler = Profiler(summarize_every=10, disabled=False)
+    for step, speaker_batch in enumerate(loader, init_step):
+        profiler.tick("Blocking, waiting for batch (threaded)")
+        
+        # Forward pass
+        inputs = torch.from_numpy(speaker_batch.data).to(device)
+        sync(device)
+        profiler.tick("Data to %s" % device)
+        embeds = model(inputs)
+        sync(device)
+        profiler.tick("Forward pass")
+        embeds_loss = embeds.view((speakers_per_batch, utterances_per_speaker, -1)).to(loss_device)
+        loss, eer = model.loss(embeds_loss)
+        sync(loss_device)
+        profiler.tick("Loss")
+
+        # Backward pass
+        model.zero_grad()
+        loss.backward()
+        profiler.tick("Backward pass")
+        model.do_gradient_ops()
+        optimizer.step()
+        profiler.tick("Parameter update")
+        
+        # Update visualizations
+        # learning_rate = optimizer.param_groups[0]["lr"]
+        vis.update(loss.item(), eer, step)
+        
+        # Draw projections and save them to the backup folder
+        if umap_every != 0 and step % umap_every == 0:
+            print("Drawing and saving projections (step %d)" % step)
+            backup_dir.mkdir(exist_ok=True)
+            projection_fpath = backup_dir.joinpath("%s_umap_%06d.png" % (run_id, step))
+            embeds = embeds.detach().cpu().numpy()
+            vis.draw_projections(embeds, utterances_per_speaker, step, projection_fpath)
+            vis.save()
+
+        # Overwrite the latest version of the model
+        if save_every != 0 and step % save_every == 0:
+            print("Saving the model (step %d)" % step)
+            torch.save({
+                "step": step + 1,
+                "model_state": model.state_dict(),
+                "optimizer_state": optimizer.state_dict(),
+            }, state_fpath)
+            
+        # Make a backup
+        if backup_every != 0 and step % backup_every == 0:
+            print("Making a backup (step %d)" % step)
+            backup_dir.mkdir(exist_ok=True)
+            backup_fpath = backup_dir.joinpath("%s_bak_%06d.pt" % (run_id, step))
+            torch.save({
+                "step": step + 1,
+                "model_state": model.state_dict(),
+                "optimizer_state": optimizer.state_dict(),
+            }, backup_fpath)
+            
+        profiler.tick("Extras (visualizations, saving)")

encoder/visualizations.py

@@ -0,0 +1,178 @@
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataset
+from datetime import datetime
+from time import perf_counter as timer
+import matplotlib.pyplot as plt
+import numpy as np
+# import webbrowser
+import visdom
+import umap
+
+colormap = np.array([
+    [76, 255, 0],
+    [0, 127, 70],
+    [255, 0, 0],
+    [255, 217, 38],
+    [0, 135, 255],
+    [165, 0, 165],
+    [255, 167, 255],
+    [0, 255, 255],
+    [255, 96, 38],
+    [142, 76, 0],
+    [33, 0, 127],
+    [0, 0, 0],
+    [183, 183, 183],
+], dtype=np.float) / 255 
+
+
+class Visualizations:
+    def __init__(self, env_name=None, update_every=10, server="http://localhost", disabled=False):
+        # Tracking data
+        self.last_update_timestamp = timer()
+        self.update_every = update_every
+        self.step_times = []
+        self.losses = []
+        self.eers = []
+        print("Updating the visualizations every %d steps." % update_every)
+        
+        # If visdom is disabled TODO: use a better paradigm for that
+        self.disabled = disabled    
+        if self.disabled:
+            return 
+        
+        # Set the environment name
+        now = str(datetime.now().strftime("%d-%m %Hh%M"))
+        if env_name is None:
+            self.env_name = now
+        else:
+            self.env_name = "%s (%s)" % (env_name, now)
+        
+        # Connect to visdom and open the corresponding window in the browser
+        try:
+            self.vis = visdom.Visdom(server, env=self.env_name, raise_exceptions=True)
+        except ConnectionError:
+            raise Exception("No visdom server detected. Run the command \"visdom\" in your CLI to "
+                            "start it.")
+        # webbrowser.open("http://localhost:8097/env/" + self.env_name)
+        
+        # Create the windows
+        self.loss_win = None
+        self.eer_win = None
+        # self.lr_win = None
+        self.implementation_win = None
+        self.projection_win = None
+        self.implementation_string = ""
+        
+    def log_params(self):
+        if self.disabled:
+            return 
+        from encoder import params_data
+        from encoder import params_model
+        param_string = "<b>Model parameters</b>:<br>"
+        for param_name in (p for p in dir(params_model) if not p.startswith("__")):
+            value = getattr(params_model, param_name)
+            param_string += "\t%s: %s<br>" % (param_name, value)
+        param_string += "<b>Data parameters</b>:<br>"
+        for param_name in (p for p in dir(params_data) if not p.startswith("__")):
+            value = getattr(params_data, param_name)
+            param_string += "\t%s: %s<br>" % (param_name, value)
+        self.vis.text(param_string, opts={"title": "Parameters"})
+        
+    def log_dataset(self, dataset: SpeakerVerificationDataset):
+        if self.disabled:
+            return 
+        dataset_string = ""
+        dataset_string += "<b>Speakers</b>: %s\n" % len(dataset.speakers)
+        dataset_string += "\n" + dataset.get_logs()
+        dataset_string = dataset_string.replace("\n", "<br>")
+        self.vis.text(dataset_string, opts={"title": "Dataset"})
+        
+    def log_implementation(self, params):
+        if self.disabled:
+            return 
+        implementation_string = ""
+        for param, value in params.items():
+            implementation_string += "<b>%s</b>: %s\n" % (param, value)
+            implementation_string = implementation_string.replace("\n", "<br>")
+        self.implementation_string = implementation_string
+        self.implementation_win = self.vis.text(
+            implementation_string, 
+            opts={"title": "Training implementation"}
+        )
+
+    def update(self, loss, eer, step):
+        # Update the tracking data
+        now = timer()
+        self.step_times.append(1000 * (now - self.last_update_timestamp))
+        self.last_update_timestamp = now
+        self.losses.append(loss)
+        self.eers.append(eer)
+        print(".", end="")
+        
+        # Update the plots every <update_every> steps
+        if step % self.update_every != 0:
+            return
+        time_string = "Step time:  mean: %5dms  std: %5dms" % \
+                      (int(np.mean(self.step_times)), int(np.std(self.step_times)))
+        print("\nStep %6d   Loss: %.4f   EER: %.4f   %s" %
+              (step, np.mean(self.losses), np.mean(self.eers), time_string))
+        if not self.disabled:
+            self.loss_win = self.vis.line(
+                [np.mean(self.losses)],
+                [step],
+                win=self.loss_win,
+                update="append" if self.loss_win else None,
+                opts=dict(
+                    legend=["Avg. loss"],
+                    xlabel="Step",
+                    ylabel="Loss",
+                    title="Loss",
+                )
+            )
+            self.eer_win = self.vis.line(
+                [np.mean(self.eers)],
+                [step],
+                win=self.eer_win,
+                update="append" if self.eer_win else None,
+                opts=dict(
+                    legend=["Avg. EER"],
+                    xlabel="Step",
+                    ylabel="EER",
+                    title="Equal error rate"
+                )
+            )
+            if self.implementation_win is not None:
+                self.vis.text(
+                    self.implementation_string + ("<b>%s</b>" % time_string), 
+                    win=self.implementation_win,
+                    opts={"title": "Training implementation"},
+                )
+
+        # Reset the tracking
+        self.losses.clear()
+        self.eers.clear()
+        self.step_times.clear()
+        
+    def draw_projections(self, embeds, utterances_per_speaker, step, out_fpath=None,
+                         max_speakers=10):
+        max_speakers = min(max_speakers, len(colormap))
+        embeds = embeds[:max_speakers * utterances_per_speaker]
+        
+        n_speakers = len(embeds) // utterances_per_speaker
+        ground_truth = np.repeat(np.arange(n_speakers), utterances_per_speaker)
+        colors = [colormap[i] for i in ground_truth]
+        
+        reducer = umap.UMAP()
+        projected = reducer.fit_transform(embeds)
+        plt.scatter(projected[:, 0], projected[:, 1], c=colors)
+        plt.gca().set_aspect("equal", "datalim")
+        plt.title("UMAP projection (step %d)" % step)
+        if not self.disabled:
+            self.projection_win = self.vis.matplot(plt, win=self.projection_win)
+        if out_fpath is not None:
+            plt.savefig(out_fpath)
+        plt.clf()
+        
+    def save(self):
+        if not self.disabled:
+            self.vis.save([self.env_name])
+        

encoder_preprocess.py

@@ -0,0 +1,70 @@
+from encoder.preprocess import preprocess_librispeech, preprocess_voxceleb1, preprocess_voxceleb2
+from utils.argutils import print_args
+from pathlib import Path
+import argparse
+
+if __name__ == "__main__":
+    class MyFormatter(argparse.ArgumentDefaultsHelpFormatter, argparse.RawDescriptionHelpFormatter):
+        pass
+    
+    parser = argparse.ArgumentParser(
+        description="Preprocesses audio files from datasets, encodes them as mel spectrograms and "
+                    "writes them to the disk. This will allow you to train the encoder. The "
+                    "datasets required are at least one of VoxCeleb1, VoxCeleb2 and LibriSpeech. "
+                    "Ideally, you should have all three. You should extract them as they are "
+                    "after having downloaded them and put them in a same directory, e.g.:\n"
+                    "-[datasets_root]\n"
+                    "  -LibriSpeech\n"
+                    "    -train-other-500\n"
+                    "  -VoxCeleb1\n"
+                    "    -wav\n"
+                    "    -vox1_meta.csv\n"
+                    "  -VoxCeleb2\n"
+                    "    -dev",
+        formatter_class=MyFormatter
+    )
+    parser.add_argument("datasets_root", type=Path, help=\
+        "Path to the directory containing your LibriSpeech/TTS and VoxCeleb datasets.")
+    parser.add_argument("-o", "--out_dir", type=Path, default=argparse.SUPPRESS, help=\
+        "Path to the output directory that will contain the mel spectrograms. If left out, "
+        "defaults to <datasets_root>/SV2TTS/encoder/")
+    parser.add_argument("-d", "--datasets", type=str, 
+                        default="librispeech_other,voxceleb1,voxceleb2", help=\
+        "Comma-separated list of the name of the datasets you want to preprocess. Only the train "
+        "set of these datasets will be used. Possible names: librispeech_other, voxceleb1, "
+        "voxceleb2.")
+    parser.add_argument("-s", "--skip_existing", action="store_true", help=\
+        "Whether to skip existing output files with the same name. Useful if this script was "
+        "interrupted.")
+    parser.add_argument("--no_trim", action="store_true", help=\
+        "Preprocess audio without trimming silences (not recommended).")
+    args = parser.parse_args()
+
+    # Verify webrtcvad is available
+    if not args.no_trim:
+        try:
+            import webrtcvad
+        except:
+            raise ModuleNotFoundError("Package 'webrtcvad' not found. This package enables "
+                "noise removal and is recommended. Please install and try again. If installation fails, "
+                "use --no_trim to disable this error message.")
+    del args.no_trim
+
+    # Process the arguments
+    args.datasets = args.datasets.split(",")
+    if not hasattr(args, "out_dir"):
+        args.out_dir = args.datasets_root.joinpath("SV2TTS", "encoder")
+    assert args.datasets_root.exists()
+    args.out_dir.mkdir(exist_ok=True, parents=True)
+
+    # Preprocess the datasets
+    print_args(args, parser)
+    preprocess_func = {
+        "librispeech_other": preprocess_librispeech,
+        "voxceleb1": preprocess_voxceleb1,
+        "voxceleb2": preprocess_voxceleb2,
+    }
+    args = vars(args)
+    for dataset in args.pop("datasets"):
+        print("Preprocessing %s" % dataset)
+        preprocess_func[dataset](**args)

encoder_train.py

@@ -0,0 +1,47 @@
+from utils.argutils import print_args
+from encoder.train import train
+from pathlib import Path
+import argparse
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Trains the speaker encoder. You must have run encoder_preprocess.py first.",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    
+    parser.add_argument("run_id", type=str, help= \
+        "Name for this model instance. If a model state from the same run ID was previously "
+        "saved, the training will restart from there. Pass -f to overwrite saved states and "
+        "restart from scratch.")
+    parser.add_argument("clean_data_root", type=Path, help= \
+        "Path to the output directory of encoder_preprocess.py. If you left the default "
+        "output directory when preprocessing, it should be <datasets_root>/SV2TTS/encoder/.")
+    parser.add_argument("-m", "--models_dir", type=Path, default="encoder/saved_models/", help=\
+        "Path to the output directory that will contain the saved model weights, as well as "
+        "backups of those weights and plots generated during training.")
+    parser.add_argument("-v", "--vis_every", type=int, default=10, help= \
+        "Number of steps between updates of the loss and the plots.")
+    parser.add_argument("-u", "--umap_every", type=int, default=100, help= \
+        "Number of steps between updates of the umap projection. Set to 0 to never update the "
+        "projections.")
+    parser.add_argument("-s", "--save_every", type=int, default=500, help= \
+        "Number of steps between updates of the model on the disk. Set to 0 to never save the "
+        "model.")
+    parser.add_argument("-b", "--backup_every", type=int, default=7500, help= \
+        "Number of steps between backups of the model. Set to 0 to never make backups of the "
+        "model.")
+    parser.add_argument("-f", "--force_restart", action="store_true", help= \
+        "Do not load any saved model.")
+    parser.add_argument("--visdom_server", type=str, default="http://localhost")
+    parser.add_argument("--no_visdom", action="store_true", help= \
+        "Disable visdom.")
+    args = parser.parse_args()
+    
+    # Process the arguments
+    args.models_dir.mkdir(exist_ok=True)
+    
+    # Run the training
+    print_args(args, parser)
+    train(**vars(args))
+    

packages.txt

@@ -0,0 +1,5 @@
+libportaudio2
+libsndfile1
+ffmpeg
+wget
+unzip

requirements.txt

@@ -0,0 +1,19 @@
+inflect==5.3.0
+librosa==0.8.1
+matplotlib==3.5.1
+numpy
+Pillow==8.4.0
+PyQt5==5.15.6
+scikit-learn==1.0.2
+scipy==1.7.3
+sounddevice==0.4.3
+SoundFile==0.10.3.post1
+tqdm==4.62.3
+umap-learn==0.5.2
+Unidecode==1.3.2
+urllib3==1.26.7
+visdom==0.1.8.9
+webrtcvad==2.0.10
+gradio==3.17.1
+gdown
+torch

samples/README.md

@@ -0,0 +1,22 @@
+The audio files in this folder are provided for toolbox testing and
+benchmarking purposes. These are the same reference utterances
+used by the SV2TTS authors to generate the audio samples located at:
+https://google.github.io/tacotron/publications/speaker_adaptation/index.html
+
+The `p240_00000.mp3` and `p260_00000.mp3` files are compressed
+versions of audios from the VCTK corpus available at:
+https://datashare.is.ed.ac.uk/handle/10283/3443
+VCTK.txt contains the copyright notices and licensing information.
+
+The `1320_00000.mp3`, `3575_00000.mp3`, `6829_00000.mp3`
+and `8230_00000.mp3` files are compressed versions of audios
+from the LibriSpeech dataset available at: https://openslr.org/12
+For these files, the following notice applies:
+```
+LibriSpeech (c) 2014 by Vassil Panayotov
+
+LibriSpeech ASR corpus is licensed under a
+Creative Commons Attribution 4.0 International License.
+
+See <http://creativecommons.org/licenses/by/4.0/>.
+```

samples/VCTK.txt

@@ -0,0 +1,94 @@
+---------------------------------------------------------------------
+                          CSTR VCTK Corpus 
+      English Multi-speaker Corpus for CSTR Voice Cloning Toolkit 
+
+                           (Version 0.92) 
+                        RELEASE September 2019
+             The Centre for Speech Technology Research
+                      University of Edinburgh 
+                        Copyright (c) 2019 
+
+                         Junichi Yamagishi
+                       jyamagis@inf.ed.ac.uk
+---------------------------------------------------------------------
+
+Overview 
+
+This CSTR VCTK Corpus includes speech data uttered by 110 English
+speakers with various accents. Each speaker reads out about 400
+sentences, which were selected from a newspaper, the rainbow passage
+and an elicitation paragraph used for the speech accent archive.
+
+The newspaper texts were taken from Herald Glasgow, with permission
+from Herald & Times Group. Each speaker has a different set of the
+newspaper texts selected based a greedy algorithm that increases the
+contextual and phonetic coverage. The details of the text selection 
+algorithms are described in the following paper: 
+
+C. Veaux, J. Yamagishi and S. King, 
+"The voice bank corpus: Design, collection and data analysis of 
+a large regional accent speech database," 
+https://doi.org/10.1109/ICSDA.2013.6709856
+
+The rainbow passage and elicitation paragraph are the same for all
+speakers. The rainbow passage can be found at International Dialects
+of English Archive:
+(http://web.ku.edu/~idea/readings/rainbow.htm). The elicitation
+paragraph is identical to the one used for the speech accent archive
+(http://accent.gmu.edu). The details of the the speech accent archive
+can be found at
+http://www.ualberta.ca/~aacl2009/PDFs/WeinbergerKunath2009AACL.pdf
+
+All speech data was recorded using an identical recording setup: an
+omni-directional microphone (DPA 4035) and a small diaphragm condenser 
+microphone with very wide bandwidth (Sennheiser MKH 800), 96kHz 
+sampling frequency at 24 bits and in a hemi-anechoic chamber of 
+the University of Edinburgh. (However, two speakers, p280 and p315 
+had technical issues of the audio recordings using MKH 800). 
+All recordings were converted into 16 bits, were downsampled to 
+48 kHz, and were manually end-pointed.
+
+This corpus was originally aimed for HMM-based text-to-speech synthesis 
+systems, especially for speaker-adaptive HMM-based speech synthesis 
+that uses average voice models trained on multiple speakers and speaker
+adaptation technologies. This corpus is also suitable for DNN-based 
+multi-speaker text-to-speech synthesis systems and waveform modeling.
+
+COPYING 
+
+This corpus is licensed under the Creative Commons License: Attribution 4.0 International 
+http://creativecommons.org/licenses/by/4.0/legalcode 
+
+VCTK VARIANTS 
+There are several variants of the VCTK corpus: 
+Speech enhancement 
+- Noisy speech database for training speech enhancement algorithms and TTS models where we added various types of noises to VCTK artificially: http://dx.doi.org/10.7488/ds/2117
+- Reverberant speech database for training speech dereverberation algorithms and TTS models where we added various types of reverberantion to VCTK artificially http://dx.doi.org/10.7488/ds/1425
+- Noisy reverberant speech database for training speech enhancement algorithms and TTS models http://dx.doi.org/10.7488/ds/2139
+- Device Recorded VCTK where speech signals of the VCTK corpus were played back and re-recorded in office environments using relatively inexpensive consumer devices http://dx.doi.org/10.7488/ds/2316
+- The Microsoft Scalable Noisy Speech Dataset (MS-SNSD) https://github.com/microsoft/MS-SNSD
+
+ASV and anti-spoofing 
+- Spoofing and Anti-Spoofing (SAS) corpus, which is a collection of synthetic speech signals produced by nine techniques, two of which are speech synthesis, and seven are voice conversion. All of them were built using the VCTK corpus. http://dx.doi.org/10.7488/ds/252
+- Automatic Speaker Verification Spoofing and Countermeasures Challenge (ASVspoof 2015) Database. This database consists of synthetic speech signals produced by ten techniques and this has been used in the first Automatic Speaker Verification Spoofing and Countermeasures Challenge (ASVspoof 2015) http://dx.doi.org/10.7488/ds/298
+- ASVspoof 2019: The 3rd Automatic Speaker Verification Spoofing and Countermeasures Challenge database. This database has been used in the 3rd Automatic Speaker Verification Spoofing and Countermeasures Challenge (ASVspoof 2019) https://doi.org/10.7488/ds/2555
+
+
+ACKNOWLEDGEMENTS
+
+The CSTR VCTK Corpus was constructed by:
+         
+        Christophe Veaux   (University of Edinburgh)
+        Junichi Yamagishi  (University of Edinburgh)
+        Kirsten MacDonald 
+
+The research leading to these results was partly funded from EPSRC
+grants EP/I031022/1 (NST) and EP/J002526/1 (CAF), from the RSE-NSFC
+grant (61111130120), and from the JST CREST (uDialogue).
+
+Please cite this corpus as follows:
+Christophe Veaux,  Junichi Yamagishi, Kirsten MacDonald, 
+"CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit",  
+The Centre for Speech Technology Research (CSTR), 
+University of Edinburgh 
+

synthesizer/LICENSE.txt

@@ -0,0 +1,24 @@
+MIT License
+
+Original work Copyright (c) 2018 Rayhane Mama (https://github.com/Rayhane-mamah)
+Original work Copyright (c) 2019 fatchord (https://github.com/fatchord)
+Modified work Copyright (c) 2019 Corentin Jemine (https://github.com/CorentinJ)
+Modified work Copyright (c) 2020 blue-fish (https://github.com/blue-fish)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

synthesizer/__init__.py

@@ -0,0 +1 @@
+#

synthesizer/audio.py

@@ -0,0 +1,206 @@
+import librosa
+import librosa.filters
+import numpy as np
+from scipy import signal
+from scipy.io import wavfile
+import soundfile as sf
+
+
+def load_wav(path, sr):
+    return librosa.core.load(path, sr=sr)[0]
+
+def save_wav(wav, path, sr):
+    wav *= 32767 / max(0.01, np.max(np.abs(wav)))
+    #proposed by @dsmiller
+    wavfile.write(path, sr, wav.astype(np.int16))
+
+def save_wavenet_wav(wav, path, sr):
+    sf.write(path, wav.astype(np.float32), sr)
+
+def preemphasis(wav, k, preemphasize=True):
+    if preemphasize:
+        return signal.lfilter([1, -k], [1], wav)
+    return wav
+
+def inv_preemphasis(wav, k, inv_preemphasize=True):
+    if inv_preemphasize:
+        return signal.lfilter([1], [1, -k], wav)
+    return wav
+
+#From https://github.com/r9y9/wavenet_vocoder/blob/master/audio.py
+def start_and_end_indices(quantized, silence_threshold=2):
+    for start in range(quantized.size):
+        if abs(quantized[start] - 127) > silence_threshold:
+            break
+    for end in range(quantized.size - 1, 1, -1):
+        if abs(quantized[end] - 127) > silence_threshold:
+            break
+    
+    assert abs(quantized[start] - 127) > silence_threshold
+    assert abs(quantized[end] - 127) > silence_threshold
+    
+    return start, end
+
+def get_hop_size(hparams):
+    hop_size = hparams.hop_size
+    if hop_size is None:
+        assert hparams.frame_shift_ms is not None
+        hop_size = int(hparams.frame_shift_ms / 1000 * hparams.sample_rate)
+    return hop_size
+
+def linearspectrogram(wav, hparams):
+    D = _stft(preemphasis(wav, hparams.preemphasis, hparams.preemphasize), hparams)
+    S = _amp_to_db(np.abs(D), hparams) - hparams.ref_level_db
+    
+    if hparams.signal_normalization:
+        return _normalize(S, hparams)
+    return S
+
+def melspectrogram(wav, hparams):
+    D = _stft(preemphasis(wav, hparams.preemphasis, hparams.preemphasize), hparams)
+    S = _amp_to_db(_linear_to_mel(np.abs(D), hparams), hparams) - hparams.ref_level_db
+    
+    if hparams.signal_normalization:
+        return _normalize(S, hparams)
+    return S
+
+def inv_linear_spectrogram(linear_spectrogram, hparams):
+    """Converts linear spectrogram to waveform using librosa"""
+    if hparams.signal_normalization:
+        D = _denormalize(linear_spectrogram, hparams)
+    else:
+        D = linear_spectrogram
+    
+    S = _db_to_amp(D + hparams.ref_level_db) #Convert back to linear
+    
+    if hparams.use_lws:
+        processor = _lws_processor(hparams)
+        D = processor.run_lws(S.astype(np.float64).T ** hparams.power)
+        y = processor.istft(D).astype(np.float32)
+        return inv_preemphasis(y, hparams.preemphasis, hparams.preemphasize)
+    else:
+        return inv_preemphasis(_griffin_lim(S ** hparams.power, hparams), hparams.preemphasis, hparams.preemphasize)
+
+def inv_mel_spectrogram(mel_spectrogram, hparams):
+    """Converts mel spectrogram to waveform using librosa"""
+    if hparams.signal_normalization:
+        D = _denormalize(mel_spectrogram, hparams)
+    else:
+        D = mel_spectrogram
+    
+    S = _mel_to_linear(_db_to_amp(D + hparams.ref_level_db), hparams)  # Convert back to linear
+    
+    if hparams.use_lws:
+        processor = _lws_processor(hparams)
+        D = processor.run_lws(S.astype(np.float64).T ** hparams.power)
+        y = processor.istft(D).astype(np.float32)
+        return inv_preemphasis(y, hparams.preemphasis, hparams.preemphasize)
+    else:
+        return inv_preemphasis(_griffin_lim(S ** hparams.power, hparams), hparams.preemphasis, hparams.preemphasize)
+
+def _lws_processor(hparams):
+    import lws
+    return lws.lws(hparams.n_fft, get_hop_size(hparams), fftsize=hparams.win_size, mode="speech")
+
+def _griffin_lim(S, hparams):
+    """librosa implementation of Griffin-Lim
+    Based on https://github.com/librosa/librosa/issues/434
+    """
+    angles = np.exp(2j * np.pi * np.random.rand(*S.shape))
+    S_complex = np.abs(S).astype(np.complex)
+    y = _istft(S_complex * angles, hparams)
+    for i in range(hparams.griffin_lim_iters):
+        angles = np.exp(1j * np.angle(_stft(y, hparams)))
+        y = _istft(S_complex * angles, hparams)
+    return y
+
+def _stft(y, hparams):
+    if hparams.use_lws:
+        return _lws_processor(hparams).stft(y).T
+    else:
+        return librosa.stft(y=y, n_fft=hparams.n_fft, hop_length=get_hop_size(hparams), win_length=hparams.win_size)
+
+def _istft(y, hparams):
+    return librosa.istft(y, hop_length=get_hop_size(hparams), win_length=hparams.win_size)
+
+##########################################################
+#Those are only correct when using lws!!! (This was messing with Wavenet quality for a long time!)
+def num_frames(length, fsize, fshift):
+    """Compute number of time frames of spectrogram
+    """
+    pad = (fsize - fshift)
+    if length % fshift == 0:
+        M = (length + pad * 2 - fsize) // fshift + 1
+    else:
+        M = (length + pad * 2 - fsize) // fshift + 2
+    return M
+
+
+def pad_lr(x, fsize, fshift):
+    """Compute left and right padding
+    """
+    M = num_frames(len(x), fsize, fshift)
+    pad = (fsize - fshift)
+    T = len(x) + 2 * pad
+    r = (M - 1) * fshift + fsize - T
+    return pad, pad + r
+##########################################################
+#Librosa correct padding
+def librosa_pad_lr(x, fsize, fshift):
+    return 0, (x.shape[0] // fshift + 1) * fshift - x.shape[0]
+
+# Conversions
+_mel_basis = None
+_inv_mel_basis = None
+
+def _linear_to_mel(spectogram, hparams):
+    global _mel_basis
+    if _mel_basis is None:
+        _mel_basis = _build_mel_basis(hparams)
+    return np.dot(_mel_basis, spectogram)
+
+def _mel_to_linear(mel_spectrogram, hparams):
+    global _inv_mel_basis
+    if _inv_mel_basis is None:
+        _inv_mel_basis = np.linalg.pinv(_build_mel_basis(hparams))
+    return np.maximum(1e-10, np.dot(_inv_mel_basis, mel_spectrogram))
+
+def _build_mel_basis(hparams):
+    assert hparams.fmax <= hparams.sample_rate // 2
+    return librosa.filters.mel(hparams.sample_rate, hparams.n_fft, n_mels=hparams.num_mels,
+                               fmin=hparams.fmin, fmax=hparams.fmax)
+
+def _amp_to_db(x, hparams):
+    min_level = np.exp(hparams.min_level_db / 20 * np.log(10))
+    return 20 * np.log10(np.maximum(min_level, x))
+
+def _db_to_amp(x):
+    return np.power(10.0, (x) * 0.05)
+
+def _normalize(S, hparams):
+    if hparams.allow_clipping_in_normalization:
+        if hparams.symmetric_mels:
+            return np.clip((2 * hparams.max_abs_value) * ((S - hparams.min_level_db) / (-hparams.min_level_db)) - hparams.max_abs_value,
+                           -hparams.max_abs_value, hparams.max_abs_value)
+        else:
+            return np.clip(hparams.max_abs_value * ((S - hparams.min_level_db) / (-hparams.min_level_db)), 0, hparams.max_abs_value)
+    
+    assert S.max() <= 0 and S.min() - hparams.min_level_db >= 0
+    if hparams.symmetric_mels:
+        return (2 * hparams.max_abs_value) * ((S - hparams.min_level_db) / (-hparams.min_level_db)) - hparams.max_abs_value
+    else:
+        return hparams.max_abs_value * ((S - hparams.min_level_db) / (-hparams.min_level_db))
+
+def _denormalize(D, hparams):
+    if hparams.allow_clipping_in_normalization:
+        if hparams.symmetric_mels:
+            return (((np.clip(D, -hparams.max_abs_value,
+                              hparams.max_abs_value) + hparams.max_abs_value) * -hparams.min_level_db / (2 * hparams.max_abs_value))
+                    + hparams.min_level_db)
+        else:
+            return ((np.clip(D, 0, hparams.max_abs_value) * -hparams.min_level_db / hparams.max_abs_value) + hparams.min_level_db)
+    
+    if hparams.symmetric_mels:
+        return (((D + hparams.max_abs_value) * -hparams.min_level_db / (2 * hparams.max_abs_value)) + hparams.min_level_db)
+    else:
+        return ((D * -hparams.min_level_db / hparams.max_abs_value) + hparams.min_level_db)

synthesizer/hparams.py

@@ -0,0 +1,92 @@
+import ast
+import pprint
+
+class HParams(object):
+    def __init__(self, **kwargs): self.__dict__.update(kwargs)
+    def __setitem__(self, key, value): setattr(self, key, value)
+    def __getitem__(self, key): return getattr(self, key)
+    def __repr__(self): return pprint.pformat(self.__dict__)
+
+    def parse(self, string):
+        # Overrides hparams from a comma-separated string of name=value pairs
+        if len(string) > 0:
+            overrides = [s.split("=") for s in string.split(",")]
+            keys, values = zip(*overrides)
+            keys = list(map(str.strip, keys))
+            values = list(map(str.strip, values))
+            for k in keys:
+                self.__dict__[k] = ast.literal_eval(values[keys.index(k)])
+        return self
+
+hparams = HParams(
+        ### Signal Processing (used in both synthesizer and vocoder)
+        sample_rate = 16000,
+        n_fft = 800,
+        num_mels = 80,
+        hop_size = 200,                             # Tacotron uses 12.5 ms frame shift (set to sample_rate * 0.0125)
+        win_size = 800,                             # Tacotron uses 50 ms frame length (set to sample_rate * 0.050)
+        fmin = 55,
+        min_level_db = -100,
+        ref_level_db = 20,
+        max_abs_value = 4.,                         # Gradient explodes if too big, premature convergence if too small.
+        preemphasis = 0.97,                         # Filter coefficient to use if preemphasize is True
+        preemphasize = True,
+
+        ### Tacotron Text-to-Speech (TTS)
+        tts_embed_dims = 512,                       # Embedding dimension for the graphemes/phoneme inputs
+        tts_encoder_dims = 256,
+        tts_decoder_dims = 128,
+        tts_postnet_dims = 512,
+        tts_encoder_K = 5,
+        tts_lstm_dims = 1024,
+        tts_postnet_K = 5,
+        tts_num_highways = 4,
+        tts_dropout = 0.5,
+        tts_cleaner_names = ["english_cleaners"],
+        tts_stop_threshold = -3.4,                  # Value below which audio generation ends.
+                                                    # For example, for a range of [-4, 4], this
+                                                    # will terminate the sequence at the first
+                                                    # frame that has all values < -3.4
+
+        ### Tacotron Training
+        tts_schedule = [(2,  1e-3,  20_000,  12),   # Progressive training schedule
+                        (2,  5e-4,  40_000,  12),   # (r, lr, step, batch_size)
+                        (2,  2e-4,  80_000,  12),   #
+                        (2,  1e-4, 160_000,  12),   # r = reduction factor (# of mel frames
+                        (2,  3e-5, 320_000,  12),   #     synthesized for each decoder iteration)
+                        (2,  1e-5, 640_000,  12)],  # lr = learning rate
+
+        tts_clip_grad_norm = 1.0,                   # clips the gradient norm to prevent explosion - set to None if not needed
+        tts_eval_interval = 500,                    # Number of steps between model evaluation (sample generation)
+                                                    # Set to -1 to generate after completing epoch, or 0 to disable
+
+        tts_eval_num_samples = 1,                   # Makes this number of samples
+
+        ### Data Preprocessing
+        max_mel_frames = 900,
+        rescale = True,
+        rescaling_max = 0.9,
+        synthesis_batch_size = 16,                  # For vocoder preprocessing and inference.
+
+        ### Mel Visualization and Griffin-Lim
+        signal_normalization = True,
+        power = 1.5,
+        griffin_lim_iters = 60,
+
+        ### Audio processing options
+        fmax = 7600,                                # Should not exceed (sample_rate // 2)
+        allow_clipping_in_normalization = True,     # Used when signal_normalization = True
+        clip_mels_length = True,                    # If true, discards samples exceeding max_mel_frames
+        use_lws = False,                            # "Fast spectrogram phase recovery using local weighted sums"
+        symmetric_mels = True,                      # Sets mel range to [-max_abs_value, max_abs_value] if True,
+                                                    #               and [0, max_abs_value] if False
+        trim_silence = True,                        # Use with sample_rate of 16000 for best results
+
+        ### SV2TTS
+        speaker_embedding_size = 256,               # Dimension for the speaker embedding
+        silence_min_duration_split = 0.4,           # Duration in seconds of a silence for an utterance to be split
+        utterance_min_duration = 1.6,               # Duration in seconds below which utterances are discarded
+        )
+
+def hparams_debug_string():
+    return str(hparams)

synthesizer/inference.py

@@ -0,0 +1,171 @@
+import torch
+from synthesizer import audio
+from synthesizer.hparams import hparams
+from synthesizer.models.tacotron import Tacotron
+from synthesizer.utils.symbols import symbols
+from synthesizer.utils.text import text_to_sequence
+from vocoder.display import simple_table
+from pathlib import Path
+from typing import Union, List
+import numpy as np
+import librosa
+
+
+class Synthesizer:
+    sample_rate = hparams.sample_rate
+    hparams = hparams
+    
+    def __init__(self, model_fpath: Path, verbose=True):
+        """
+        The model isn't instantiated and loaded in memory until needed or until load() is called.
+        
+        :param model_fpath: path to the trained model file
+        :param verbose: if False, prints less information when using the model
+        """
+        self.model_fpath = model_fpath
+        self.verbose = verbose
+ 
+        # Check for GPU
+        if torch.cuda.is_available():
+            self.device = torch.device("cuda")
+        else:
+            self.device = torch.device("cpu")
+        if self.verbose:
+            print("Synthesizer using device:", self.device)
+        
+        # Tacotron model will be instantiated later on first use.
+        self._model = None
+
+    def is_loaded(self):
+        """
+        Whether the model is loaded in memory.
+        """
+        return self._model is not None
+    
+    def load(self):
+        """
+        Instantiates and loads the model given the weights file that was passed in the constructor.
+        """
+        self._model = Tacotron(embed_dims=hparams.tts_embed_dims,
+                               num_chars=len(symbols),
+                               encoder_dims=hparams.tts_encoder_dims,
+                               decoder_dims=hparams.tts_decoder_dims,
+                               n_mels=hparams.num_mels,
+                               fft_bins=hparams.num_mels,
+                               postnet_dims=hparams.tts_postnet_dims,
+                               encoder_K=hparams.tts_encoder_K,
+                               lstm_dims=hparams.tts_lstm_dims,
+                               postnet_K=hparams.tts_postnet_K,
+                               num_highways=hparams.tts_num_highways,
+                               dropout=hparams.tts_dropout,
+                               stop_threshold=hparams.tts_stop_threshold,
+                               speaker_embedding_size=hparams.speaker_embedding_size).to(self.device)
+
+        self._model.load(self.model_fpath)
+        self._model.eval()
+
+        if self.verbose:
+            print("Loaded synthesizer \"%s\" trained to step %d" % (self.model_fpath.name, self._model.state_dict()["step"]))
+
+    def synthesize_spectrograms(self, texts: List[str],
+                                embeddings: Union[np.ndarray, List[np.ndarray]],
+                                return_alignments=False):
+        """
+        Synthesizes mel spectrograms from texts and speaker embeddings.
+
+        :param texts: a list of N text prompts to be synthesized
+        :param embeddings: a numpy array or list of speaker embeddings of shape (N, 256) 
+        :param return_alignments: if True, a matrix representing the alignments between the 
+        characters
+        and each decoder output step will be returned for each spectrogram
+        :return: a list of N melspectrograms as numpy arrays of shape (80, Mi), where Mi is the 
+        sequence length of spectrogram i, and possibly the alignments.
+        """
+        # Load the model on the first request.
+        if not self.is_loaded():
+            self.load()
+
+            # Print some info about the model when it is loaded            
+            tts_k = self._model.get_step() // 1000
+
+            simple_table([("Tacotron", str(tts_k) + "k"),
+                        ("r", self._model.r)])
+
+        # Preprocess text inputs
+        inputs = [text_to_sequence(text.strip(), hparams.tts_cleaner_names) for text in texts]
+        if not isinstance(embeddings, list):
+            embeddings = [embeddings]
+
+        # Batch inputs
+        batched_inputs = [inputs[i:i+hparams.synthesis_batch_size]
+                             for i in range(0, len(inputs), hparams.synthesis_batch_size)]
+        batched_embeds = [embeddings[i:i+hparams.synthesis_batch_size]
+                             for i in range(0, len(embeddings), hparams.synthesis_batch_size)]
+
+        specs = []
+        for i, batch in enumerate(batched_inputs, 1):
+            if self.verbose:
+                print(f"\n| Generating {i}/{len(batched_inputs)}")
+
+            # Pad texts so they are all the same length
+            text_lens = [len(text) for text in batch]
+            max_text_len = max(text_lens)
+            chars = [pad1d(text, max_text_len) for text in batch]
+            chars = np.stack(chars)
+
+            # Stack speaker embeddings into 2D array for batch processing
+            speaker_embeds = np.stack(batched_embeds[i-1])
+
+            # Convert to tensor
+            chars = torch.tensor(chars).long().to(self.device)
+            speaker_embeddings = torch.tensor(speaker_embeds).float().to(self.device)
+
+            # Inference
+            _, mels, alignments = self._model.generate(chars, speaker_embeddings)
+            mels = mels.detach().cpu().numpy()
+            for m in mels:
+                # Trim silence from end of each spectrogram
+                while np.max(m[:, -1]) < hparams.tts_stop_threshold:
+                    m = m[:, :-1]
+                specs.append(m)
+
+        if self.verbose:
+            print("\n\nDone.\n")
+        return (specs, alignments) if return_alignments else specs
+
+    @staticmethod
+    def load_preprocess_wav(fpath):
+        """
+        Loads and preprocesses an audio file under the same conditions the audio files were used to
+        train the synthesizer. 
+        """
+        wav = librosa.load(str(fpath), hparams.sample_rate)[0]
+        if hparams.rescale:
+            wav = wav / np.abs(wav).max() * hparams.rescaling_max
+        return wav
+
+    @staticmethod
+    def make_spectrogram(fpath_or_wav: Union[str, Path, np.ndarray]):
+        """
+        Creates a mel spectrogram from an audio file in the same manner as the mel spectrograms that 
+        were fed to the synthesizer when training.
+        """
+        if isinstance(fpath_or_wav, str) or isinstance(fpath_or_wav, Path):
+            wav = Synthesizer.load_preprocess_wav(fpath_or_wav)
+        else:
+            wav = fpath_or_wav
+        
+        mel_spectrogram = audio.melspectrogram(wav, hparams).astype(np.float32)
+        return mel_spectrogram
+    
+    @staticmethod
+    def griffin_lim(mel):
+        """
+        Inverts a mel spectrogram using Griffin-Lim. The mel spectrogram is expected to have been built
+        with the same parameters present in hparams.py.
+        """
+        return audio.inv_mel_spectrogram(mel, hparams)
+
+
+def pad1d(x, max_len, pad_value=0):
+    return np.pad(x, (0, max_len - len(x)), mode="constant", constant_values=pad_value)

synthesizer/models/tacotron.py

@@ -0,0 +1,519 @@
+import os
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from pathlib import Path
+from typing import Union
+
+
+class HighwayNetwork(nn.Module):
+    def __init__(self, size):
+        super().__init__()
+        self.W1 = nn.Linear(size, size)
+        self.W2 = nn.Linear(size, size)
+        self.W1.bias.data.fill_(0.)
+
+    def forward(self, x):
+        x1 = self.W1(x)
+        x2 = self.W2(x)
+        g = torch.sigmoid(x2)
+        y = g * F.relu(x1) + (1. - g) * x
+        return y
+
+
+class Encoder(nn.Module):
+    def __init__(self, embed_dims, num_chars, encoder_dims, K, num_highways, dropout):
+        super().__init__()
+        prenet_dims = (encoder_dims, encoder_dims)
+        cbhg_channels = encoder_dims
+        self.embedding = nn.Embedding(num_chars, embed_dims)
+        self.pre_net = PreNet(embed_dims, fc1_dims=prenet_dims[0], fc2_dims=prenet_dims[1],
+                              dropout=dropout)
+        self.cbhg = CBHG(K=K, in_channels=cbhg_channels, channels=cbhg_channels,
+                         proj_channels=[cbhg_channels, cbhg_channels],
+                         num_highways=num_highways)
+
+    def forward(self, x, speaker_embedding=None):
+        x = self.embedding(x)
+        x = self.pre_net(x)
+        x.transpose_(1, 2)
+        x = self.cbhg(x)
+        if speaker_embedding is not None:
+            x = self.add_speaker_embedding(x, speaker_embedding)
+        return x
+
+    def add_speaker_embedding(self, x, speaker_embedding):
+        # SV2TTS
+        # The input x is the encoder output and is a 3D tensor with size (batch_size, num_chars, tts_embed_dims)
+        # When training, speaker_embedding is also a 2D tensor with size (batch_size, speaker_embedding_size)
+        #     (for inference, speaker_embedding is a 1D tensor with size (speaker_embedding_size))
+        # This concats the speaker embedding for each char in the encoder output
+
+        # Save the dimensions as human-readable names
+        batch_size = x.size()[0]
+        num_chars = x.size()[1]
+
+        if speaker_embedding.dim() == 1:
+            idx = 0
+        else:
+            idx = 1
+
+        # Start by making a copy of each speaker embedding to match the input text length
+        # The output of this has size (batch_size, num_chars * tts_embed_dims)
+        speaker_embedding_size = speaker_embedding.size()[idx]
+        e = speaker_embedding.repeat_interleave(num_chars, dim=idx)
+
+        # Reshape it and transpose
+        e = e.reshape(batch_size, speaker_embedding_size, num_chars)
+        e = e.transpose(1, 2)
+
+        # Concatenate the tiled speaker embedding with the encoder output
+        x = torch.cat((x, e), 2)
+        return x
+
+
+class BatchNormConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel, relu=True):
+        super().__init__()
+        self.conv = nn.Conv1d(in_channels, out_channels, kernel, stride=1, padding=kernel // 2, bias=False)
+        self.bnorm = nn.BatchNorm1d(out_channels)
+        self.relu = relu
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = F.relu(x) if self.relu is True else x
+        return self.bnorm(x)
+
+
+class CBHG(nn.Module):
+    def __init__(self, K, in_channels, channels, proj_channels, num_highways):
+        super().__init__()
+
+        # List of all rnns to call `flatten_parameters()` on
+        self._to_flatten = []
+
+        self.bank_kernels = [i for i in range(1, K + 1)]
+        self.conv1d_bank = nn.ModuleList()
+        for k in self.bank_kernels:
+            conv = BatchNormConv(in_channels, channels, k)
+            self.conv1d_bank.append(conv)
+
+        self.maxpool = nn.MaxPool1d(kernel_size=2, stride=1, padding=1)
+
+        self.conv_project1 = BatchNormConv(len(self.bank_kernels) * channels, proj_channels[0], 3)
+        self.conv_project2 = BatchNormConv(proj_channels[0], proj_channels[1], 3, relu=False)
+
+        # Fix the highway input if necessary
+        if proj_channels[-1] != channels:
+            self.highway_mismatch = True
+            self.pre_highway = nn.Linear(proj_channels[-1], channels, bias=False)
+        else:
+            self.highway_mismatch = False
+
+        self.highways = nn.ModuleList()
+        for i in range(num_highways):
+            hn = HighwayNetwork(channels)
+            self.highways.append(hn)
+
+        self.rnn = nn.GRU(channels, channels // 2, batch_first=True, bidirectional=True)
+        self._to_flatten.append(self.rnn)
+
+        # Avoid fragmentation of RNN parameters and associated warning
+        self._flatten_parameters()
+
+    def forward(self, x):
+        # Although we `_flatten_parameters()` on init, when using DataParallel
+        # the model gets replicated, making it no longer guaranteed that the
+        # weights are contiguous in GPU memory. Hence, we must call it again
+        self._flatten_parameters()
+
+        # Save these for later
+        residual = x
+        seq_len = x.size(-1)
+        conv_bank = []
+
+        # Convolution Bank
+        for conv in self.conv1d_bank:
+            c = conv(x) # Convolution
+            conv_bank.append(c[:, :, :seq_len])
+
+        # Stack along the channel axis
+        conv_bank = torch.cat(conv_bank, dim=1)
+
+        # dump the last padding to fit residual
+        x = self.maxpool(conv_bank)[:, :, :seq_len]
+
+        # Conv1d projections
+        x = self.conv_project1(x)
+        x = self.conv_project2(x)
+
+        # Residual Connect
+        x = x + residual
+
+        # Through the highways
+        x = x.transpose(1, 2)
+        if self.highway_mismatch is True:
+            x = self.pre_highway(x)
+        for h in self.highways: x = h(x)
+
+        # And then the RNN
+        x, _ = self.rnn(x)
+        return x
+
+    def _flatten_parameters(self):
+        """Calls `flatten_parameters` on all the rnns used by the WaveRNN. Used
+        to improve efficiency and avoid PyTorch yelling at us."""
+        [m.flatten_parameters() for m in self._to_flatten]
+
+class PreNet(nn.Module):
+    def __init__(self, in_dims, fc1_dims=256, fc2_dims=128, dropout=0.5):
+        super().__init__()
+        self.fc1 = nn.Linear(in_dims, fc1_dims)
+        self.fc2 = nn.Linear(fc1_dims, fc2_dims)
+        self.p = dropout
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = F.relu(x)
+        x = F.dropout(x, self.p, training=True)
+        x = self.fc2(x)
+        x = F.relu(x)
+        x = F.dropout(x, self.p, training=True)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, attn_dims):
+        super().__init__()
+        self.W = nn.Linear(attn_dims, attn_dims, bias=False)
+        self.v = nn.Linear(attn_dims, 1, bias=False)
+
+    def forward(self, encoder_seq_proj, query, t):
+
+        # print(encoder_seq_proj.shape)
+        # Transform the query vector
+        query_proj = self.W(query).unsqueeze(1)
+
+        # Compute the scores
+        u = self.v(torch.tanh(encoder_seq_proj + query_proj))
+        scores = F.softmax(u, dim=1)
+
+        return scores.transpose(1, 2)
+
+
+class LSA(nn.Module):
+    def __init__(self, attn_dim, kernel_size=31, filters=32):
+        super().__init__()
+        self.conv = nn.Conv1d(1, filters, padding=(kernel_size - 1) // 2, kernel_size=kernel_size, bias=True)
+        self.L = nn.Linear(filters, attn_dim, bias=False)
+        self.W = nn.Linear(attn_dim, attn_dim, bias=True) # Include the attention bias in this term
+        self.v = nn.Linear(attn_dim, 1, bias=False)
+        self.cumulative = None
+        self.attention = None
+
+    def init_attention(self, encoder_seq_proj):
+        device = next(self.parameters()).device  # use same device as parameters
+        b, t, c = encoder_seq_proj.size()
+        self.cumulative = torch.zeros(b, t, device=device)
+        self.attention = torch.zeros(b, t, device=device)
+
+    def forward(self, encoder_seq_proj, query, t, chars):
+
+        if t == 0: self.init_attention(encoder_seq_proj)
+
+        processed_query = self.W(query).unsqueeze(1)
+
+        location = self.cumulative.unsqueeze(1)
+        processed_loc = self.L(self.conv(location).transpose(1, 2))
+
+        u = self.v(torch.tanh(processed_query + encoder_seq_proj + processed_loc))
+        u = u.squeeze(-1)
+
+        # Mask zero padding chars
+        u = u * (chars != 0).float()
+
+        # Smooth Attention
+        # scores = torch.sigmoid(u) / torch.sigmoid(u).sum(dim=1, keepdim=True)
+        scores = F.softmax(u, dim=1)
+        self.attention = scores
+        self.cumulative = self.cumulative + self.attention
+
+        return scores.unsqueeze(-1).transpose(1, 2)
+
+
+class Decoder(nn.Module):
+    # Class variable because its value doesn't change between classes
+    # yet ought to be scoped by class because its a property of a Decoder
+    max_r = 20
+    def __init__(self, n_mels, encoder_dims, decoder_dims, lstm_dims,
+                 dropout, speaker_embedding_size):
+        super().__init__()
+        self.register_buffer("r", torch.tensor(1, dtype=torch.int))
+        self.n_mels = n_mels
+        prenet_dims = (decoder_dims * 2, decoder_dims * 2)
+        self.prenet = PreNet(n_mels, fc1_dims=prenet_dims[0], fc2_dims=prenet_dims[1],
+                             dropout=dropout)
+        self.attn_net = LSA(decoder_dims)
+        self.attn_rnn = nn.GRUCell(encoder_dims + prenet_dims[1] + speaker_embedding_size, decoder_dims)
+        self.rnn_input = nn.Linear(encoder_dims + decoder_dims + speaker_embedding_size, lstm_dims)
+        self.res_rnn1 = nn.LSTMCell(lstm_dims, lstm_dims)
+        self.res_rnn2 = nn.LSTMCell(lstm_dims, lstm_dims)
+        self.mel_proj = nn.Linear(lstm_dims, n_mels * self.max_r, bias=False)
+        self.stop_proj = nn.Linear(encoder_dims + speaker_embedding_size + lstm_dims, 1)
+
+    def zoneout(self, prev, current, p=0.1):
+        device = next(self.parameters()).device  # Use same device as parameters
+        mask = torch.zeros(prev.size(), device=device).bernoulli_(p)
+        return prev * mask + current * (1 - mask)
+
+    def forward(self, encoder_seq, encoder_seq_proj, prenet_in,
+                hidden_states, cell_states, context_vec, t, chars):
+
+        # Need this for reshaping mels
+        batch_size = encoder_seq.size(0)
+
+        # Unpack the hidden and cell states
+        attn_hidden, rnn1_hidden, rnn2_hidden = hidden_states
+        rnn1_cell, rnn2_cell = cell_states
+
+        # PreNet for the Attention RNN
+        prenet_out = self.prenet(prenet_in)
+
+        # Compute the Attention RNN hidden state
+        attn_rnn_in = torch.cat([context_vec, prenet_out], dim=-1)
+        attn_hidden = self.attn_rnn(attn_rnn_in.squeeze(1), attn_hidden)
+
+        # Compute the attention scores
+        scores = self.attn_net(encoder_seq_proj, attn_hidden, t, chars)
+
+        # Dot product to create the context vector
+        context_vec = scores @ encoder_seq
+        context_vec = context_vec.squeeze(1)
+
+        # Concat Attention RNN output w. Context Vector & project
+        x = torch.cat([context_vec, attn_hidden], dim=1)
+        x = self.rnn_input(x)
+
+        # Compute first Residual RNN
+        rnn1_hidden_next, rnn1_cell = self.res_rnn1(x, (rnn1_hidden, rnn1_cell))
+        if self.training:
+            rnn1_hidden = self.zoneout(rnn1_hidden, rnn1_hidden_next)
+        else:
+            rnn1_hidden = rnn1_hidden_next
+        x = x + rnn1_hidden
+
+        # Compute second Residual RNN
+        rnn2_hidden_next, rnn2_cell = self.res_rnn2(x, (rnn2_hidden, rnn2_cell))
+        if self.training:
+            rnn2_hidden = self.zoneout(rnn2_hidden, rnn2_hidden_next)
+        else:
+            rnn2_hidden = rnn2_hidden_next
+        x = x + rnn2_hidden
+
+        # Project Mels
+        mels = self.mel_proj(x)
+        mels = mels.view(batch_size, self.n_mels, self.max_r)[:, :, :self.r]
+        hidden_states = (attn_hidden, rnn1_hidden, rnn2_hidden)
+        cell_states = (rnn1_cell, rnn2_cell)
+
+        # Stop token prediction
+        s = torch.cat((x, context_vec), dim=1)
+        s = self.stop_proj(s)
+        stop_tokens = torch.sigmoid(s)
+
+        return mels, scores, hidden_states, cell_states, context_vec, stop_tokens
+
+
+class Tacotron(nn.Module):
+    def __init__(self, embed_dims, num_chars, encoder_dims, decoder_dims, n_mels, 
+                 fft_bins, postnet_dims, encoder_K, lstm_dims, postnet_K, num_highways,
+                 dropout, stop_threshold, speaker_embedding_size):
+        super().__init__()
+        self.n_mels = n_mels
+        self.lstm_dims = lstm_dims
+        self.encoder_dims = encoder_dims
+        self.decoder_dims = decoder_dims
+        self.speaker_embedding_size = speaker_embedding_size
+        self.encoder = Encoder(embed_dims, num_chars, encoder_dims,
+                               encoder_K, num_highways, dropout)
+        self.encoder_proj = nn.Linear(encoder_dims + speaker_embedding_size, decoder_dims, bias=False)
+        self.decoder = Decoder(n_mels, encoder_dims, decoder_dims, lstm_dims,
+                               dropout, speaker_embedding_size)
+        self.postnet = CBHG(postnet_K, n_mels, postnet_dims,
+                            [postnet_dims, fft_bins], num_highways)
+        self.post_proj = nn.Linear(postnet_dims, fft_bins, bias=False)
+
+        self.init_model()
+        self.num_params()
+
+        self.register_buffer("step", torch.zeros(1, dtype=torch.long))
+        self.register_buffer("stop_threshold", torch.tensor(stop_threshold, dtype=torch.float32))
+
+    @property
+    def r(self):
+        return self.decoder.r.item()
+
+    @r.setter
+    def r(self, value):
+        self.decoder.r = self.decoder.r.new_tensor(value, requires_grad=False)
+
+    def forward(self, x, m, speaker_embedding):
+        device = next(self.parameters()).device  # use same device as parameters
+
+        self.step += 1
+        batch_size, _, steps  = m.size()
+
+        # Initialise all hidden states and pack into tuple
+        attn_hidden = torch.zeros(batch_size, self.decoder_dims, device=device)
+        rnn1_hidden = torch.zeros(batch_size, self.lstm_dims, device=device)
+        rnn2_hidden = torch.zeros(batch_size, self.lstm_dims, device=device)
+        hidden_states = (attn_hidden, rnn1_hidden, rnn2_hidden)
+
+        # Initialise all lstm cell states and pack into tuple
+        rnn1_cell = torch.zeros(batch_size, self.lstm_dims, device=device)
+        rnn2_cell = torch.zeros(batch_size, self.lstm_dims, device=device)
+        cell_states = (rnn1_cell, rnn2_cell)
+
+        # <GO> Frame for start of decoder loop
+        go_frame = torch.zeros(batch_size, self.n_mels, device=device)
+
+        # Need an initial context vector
+        context_vec = torch.zeros(batch_size, self.encoder_dims + self.speaker_embedding_size, device=device)
+
+        # SV2TTS: Run the encoder with the speaker embedding
+        # The projection avoids unnecessary matmuls in the decoder loop
+        encoder_seq = self.encoder(x, speaker_embedding)
+        encoder_seq_proj = self.encoder_proj(encoder_seq)
+
+        # Need a couple of lists for outputs
+        mel_outputs, attn_scores, stop_outputs = [], [], []
+
+        # Run the decoder loop
+        for t in range(0, steps, self.r):
+            prenet_in = m[:, :, t - 1] if t > 0 else go_frame
+            mel_frames, scores, hidden_states, cell_states, context_vec, stop_tokens = \
+                self.decoder(encoder_seq, encoder_seq_proj, prenet_in,
+                             hidden_states, cell_states, context_vec, t, x)
+            mel_outputs.append(mel_frames)
+            attn_scores.append(scores)
+            stop_outputs.extend([stop_tokens] * self.r)
+
+        # Concat the mel outputs into sequence
+        mel_outputs = torch.cat(mel_outputs, dim=2)
+
+        # Post-Process for Linear Spectrograms
+        postnet_out = self.postnet(mel_outputs)
+        linear = self.post_proj(postnet_out)
+        linear = linear.transpose(1, 2)
+
+        # For easy visualisation
+        attn_scores = torch.cat(attn_scores, 1)
+        # attn_scores = attn_scores.cpu().data.numpy()
+        stop_outputs = torch.cat(stop_outputs, 1)
+
+        return mel_outputs, linear, attn_scores, stop_outputs
+
+    def generate(self, x, speaker_embedding=None, steps=2000):
+        self.eval()
+        device = next(self.parameters()).device  # use same device as parameters
+
+        batch_size, _  = x.size()
+
+        # Need to initialise all hidden states and pack into tuple for tidyness
+        attn_hidden = torch.zeros(batch_size, self.decoder_dims, device=device)
+        rnn1_hidden = torch.zeros(batch_size, self.lstm_dims, device=device)
+        rnn2_hidden = torch.zeros(batch_size, self.lstm_dims, device=device)
+        hidden_states = (attn_hidden, rnn1_hidden, rnn2_hidden)
+
+        # Need to initialise all lstm cell states and pack into tuple for tidyness
+        rnn1_cell = torch.zeros(batch_size, self.lstm_dims, device=device)
+        rnn2_cell = torch.zeros(batch_size, self.lstm_dims, device=device)
+        cell_states = (rnn1_cell, rnn2_cell)
+
+        # Need a <GO> Frame for start of decoder loop
+        go_frame = torch.zeros(batch_size, self.n_mels, device=device)
+
+        # Need an initial context vector
+        context_vec = torch.zeros(batch_size, self.encoder_dims + self.speaker_embedding_size, device=device)
+
+        # SV2TTS: Run the encoder with the speaker embedding
+        # The projection avoids unnecessary matmuls in the decoder loop
+        encoder_seq = self.encoder(x, speaker_embedding)
+        encoder_seq_proj = self.encoder_proj(encoder_seq)
+
+        # Need a couple of lists for outputs
+        mel_outputs, attn_scores, stop_outputs = [], [], []
+
+        # Run the decoder loop
+        for t in range(0, steps, self.r):
+            prenet_in = mel_outputs[-1][:, :, -1] if t > 0 else go_frame
+            mel_frames, scores, hidden_states, cell_states, context_vec, stop_tokens = \
+            self.decoder(encoder_seq, encoder_seq_proj, prenet_in,
+                         hidden_states, cell_states, context_vec, t, x)
+            mel_outputs.append(mel_frames)
+            attn_scores.append(scores)
+            stop_outputs.extend([stop_tokens] * self.r)
+            # Stop the loop when all stop tokens in batch exceed threshold
+            if (stop_tokens > 0.5).all() and t > 10: break
+
+        # Concat the mel outputs into sequence
+        mel_outputs = torch.cat(mel_outputs, dim=2)
+
+        # Post-Process for Linear Spectrograms
+        postnet_out = self.postnet(mel_outputs)
+        linear = self.post_proj(postnet_out)
+
+
+        linear = linear.transpose(1, 2)
+
+        # For easy visualisation
+        attn_scores = torch.cat(attn_scores, 1)
+        stop_outputs = torch.cat(stop_outputs, 1)
+
+        self.train()
+
+        return mel_outputs, linear, attn_scores
+
+    def init_model(self):
+        for p in self.parameters():
+            if p.dim() > 1: nn.init.xavier_uniform_(p)
+
+    def get_step(self):
+        return self.step.data.item()
+
+    def reset_step(self):
+        # assignment to parameters or buffers is overloaded, updates internal dict entry
+        self.step = self.step.data.new_tensor(1)
+
+    def log(self, path, msg):
+        with open(path, "a") as f:
+            print(msg, file=f)
+
+    def load(self, path, optimizer=None):
+        # Use device of model params as location for loaded state
+        device = next(self.parameters()).device
+        checkpoint = torch.load(str(path), map_location=device)
+        self.load_state_dict(checkpoint["model_state"])
+
+        if "optimizer_state" in checkpoint and optimizer is not None:
+            optimizer.load_state_dict(checkpoint["optimizer_state"])
+
+    def save(self, path, optimizer=None):
+        if optimizer is not None:
+            torch.save({
+                "model_state": self.state_dict(),
+                "optimizer_state": optimizer.state_dict(),
+            }, str(path))
+        else:
+            torch.save({
+                "model_state": self.state_dict(),
+            }, str(path))
+
+
+    def num_params(self, print_out=True):
+        parameters = filter(lambda p: p.requires_grad, self.parameters())
+        parameters = sum([np.prod(p.size()) for p in parameters]) / 1_000_000
+        if print_out:
+            print("Trainable Parameters: %.3fM" % parameters)
+        return parameters

synthesizer/preprocess.py

@@ -0,0 +1,259 @@
+from multiprocessing.pool import Pool 
+from synthesizer import audio
+from functools import partial
+from itertools import chain
+from encoder import inference as encoder
+from pathlib import Path
+from utils import logmmse
+from tqdm import tqdm
+import numpy as np
+import librosa
+
+
+def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
+                           skip_existing: bool, hparams, no_alignments: bool,
+                           datasets_name: str, subfolders: str):
+    # Gather the input directories
+    dataset_root = datasets_root.joinpath(datasets_name)
+    input_dirs = [dataset_root.joinpath(subfolder.strip()) for subfolder in subfolders.split(",")]
+    print("\n    ".join(map(str, ["Using data from:"] + input_dirs)))
+    assert all(input_dir.exists() for input_dir in input_dirs)
+    
+    # Create the output directories for each output file type
+    out_dir.joinpath("mels").mkdir(exist_ok=True)
+    out_dir.joinpath("audio").mkdir(exist_ok=True)
+    
+    # Create a metadata file
+    metadata_fpath = out_dir.joinpath("train.txt")
+    metadata_file = metadata_fpath.open("a" if skip_existing else "w", encoding="utf-8")
+
+    # Preprocess the dataset
+    speaker_dirs = list(chain.from_iterable(input_dir.glob("*") for input_dir in input_dirs))
+    func = partial(preprocess_speaker, out_dir=out_dir, skip_existing=skip_existing, 
+                   hparams=hparams, no_alignments=no_alignments)
+    job = Pool(n_processes).imap(func, speaker_dirs)
+    for speaker_metadata in tqdm(job, datasets_name, len(speaker_dirs), unit="speakers"):
+        for metadatum in speaker_metadata:
+            metadata_file.write("|".join(str(x) for x in metadatum) + "\n")
+    metadata_file.close()
+
+    # Verify the contents of the metadata file
+    with metadata_fpath.open("r", encoding="utf-8") as metadata_file:
+        metadata = [line.split("|") for line in metadata_file]
+    mel_frames = sum([int(m[4]) for m in metadata])
+    timesteps = sum([int(m[3]) for m in metadata])
+    sample_rate = hparams.sample_rate
+    hours = (timesteps / sample_rate) / 3600
+    print("The dataset consists of %d utterances, %d mel frames, %d audio timesteps (%.2f hours)." %
+          (len(metadata), mel_frames, timesteps, hours))
+    print("Max input length (text chars): %d" % max(len(m[5]) for m in metadata))
+    print("Max mel frames length: %d" % max(int(m[4]) for m in metadata))
+    print("Max audio timesteps length: %d" % max(int(m[3]) for m in metadata))
+
+
+def preprocess_speaker(speaker_dir, out_dir: Path, skip_existing: bool, hparams, no_alignments: bool):
+    metadata = []
+    for book_dir in speaker_dir.glob("*"):
+        if no_alignments:
+            # Gather the utterance audios and texts
+            # LibriTTS uses .wav but we will include extensions for compatibility with other datasets
+            extensions = ["*.wav", "*.flac", "*.mp3"]
+            for extension in extensions:
+                wav_fpaths = book_dir.glob(extension)
+
+                for wav_fpath in wav_fpaths:
+                    # Load the audio waveform
+                    wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
+                    if hparams.rescale:
+                        wav = wav / np.abs(wav).max() * hparams.rescaling_max
+
+                    # Get the corresponding text
+                    # Check for .txt (for compatibility with other datasets)
+                    text_fpath = wav_fpath.with_suffix(".txt")
+                    if not text_fpath.exists():
+                        # Check for .normalized.txt (LibriTTS)
+                        text_fpath = wav_fpath.with_suffix(".normalized.txt")
+                        assert text_fpath.exists()
+                    with text_fpath.open("r") as text_file:
+                        text = "".join([line for line in text_file])
+                        text = text.replace("\"", "")
+                        text = text.strip()
+
+                    # Process the utterance
+                    metadata.append(process_utterance(wav, text, out_dir, str(wav_fpath.with_suffix("").name),
+                                                      skip_existing, hparams))
+        else:
+            # Process alignment file (LibriSpeech support)
+            # Gather the utterance audios and texts
+            try:
+                alignments_fpath = next(book_dir.glob("*.alignment.txt"))
+                with alignments_fpath.open("r") as alignments_file:
+                    alignments = [line.rstrip().split(" ") for line in alignments_file]
+            except StopIteration:
+                # A few alignment files will be missing
+                continue
+
+            # Iterate over each entry in the alignments file
+            for wav_fname, words, end_times in alignments:
+                wav_fpath = book_dir.joinpath(wav_fname + ".flac")
+                assert wav_fpath.exists()
+                words = words.replace("\"", "").split(",")
+                end_times = list(map(float, end_times.replace("\"", "").split(",")))
+
+                # Process each sub-utterance
+                wavs, texts = split_on_silences(wav_fpath, words, end_times, hparams)
+                for i, (wav, text) in enumerate(zip(wavs, texts)):
+                    sub_basename = "%s_%02d" % (wav_fname, i)
+                    metadata.append(process_utterance(wav, text, out_dir, sub_basename,
+                                                      skip_existing, hparams))
+
+    return [m for m in metadata if m is not None]
+
+
+def split_on_silences(wav_fpath, words, end_times, hparams):
+    # Load the audio waveform
+    wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
+    if hparams.rescale:
+        wav = wav / np.abs(wav).max() * hparams.rescaling_max
+    
+    words = np.array(words)
+    start_times = np.array([0.0] + end_times[:-1])
+    end_times = np.array(end_times)
+    assert len(words) == len(end_times) == len(start_times)
+    assert words[0] == "" and words[-1] == ""
+    
+    # Find pauses that are too long
+    mask = (words == "") & (end_times - start_times >= hparams.silence_min_duration_split)
+    mask[0] = mask[-1] = True
+    breaks = np.where(mask)[0]
+
+    # Profile the noise from the silences and perform noise reduction on the waveform
+    silence_times = [[start_times[i], end_times[i]] for i in breaks]
+    silence_times = (np.array(silence_times) * hparams.sample_rate).astype(np.int)
+    noisy_wav = np.concatenate([wav[stime[0]:stime[1]] for stime in silence_times])
+    if len(noisy_wav) > hparams.sample_rate * 0.02:
+        profile = logmmse.profile_noise(noisy_wav, hparams.sample_rate)
+        wav = logmmse.denoise(wav, profile, eta=0)
+    
+    # Re-attach segments that are too short
+    segments = list(zip(breaks[:-1], breaks[1:]))
+    segment_durations = [start_times[end] - end_times[start] for start, end in segments]
+    i = 0
+    while i < len(segments) and len(segments) > 1:
+        if segment_durations[i] < hparams.utterance_min_duration:
+            # See if the segment can be re-attached with the right or the left segment
+            left_duration = float("inf") if i == 0 else segment_durations[i - 1]
+            right_duration = float("inf") if i == len(segments) - 1 else segment_durations[i + 1]
+            joined_duration = segment_durations[i] + min(left_duration, right_duration)
+
+            # Do not re-attach if it causes the joined utterance to be too long
+            if joined_duration > hparams.hop_size * hparams.max_mel_frames / hparams.sample_rate:
+                i += 1
+                continue
+
+            # Re-attach the segment with the neighbour of shortest duration
+            j = i - 1 if left_duration <= right_duration else i
+            segments[j] = (segments[j][0], segments[j + 1][1])
+            segment_durations[j] = joined_duration
+            del segments[j + 1], segment_durations[j + 1]
+        else:
+            i += 1
+    
+    # Split the utterance
+    segment_times = [[end_times[start], start_times[end]] for start, end in segments]
+    segment_times = (np.array(segment_times) * hparams.sample_rate).astype(np.int)
+    wavs = [wav[segment_time[0]:segment_time[1]] for segment_time in segment_times]
+    texts = [" ".join(words[start + 1:end]).replace("  ", " ") for start, end in segments]
+    
+    # # DEBUG: play the audio segments (run with -n=1)
+    # import sounddevice as sd
+    # if len(wavs) > 1:
+    #     print("This sentence was split in %d segments:" % len(wavs))
+    # else:
+    #     print("There are no silences long enough for this sentence to be split:")
+    # for wav, text in zip(wavs, texts):
+    #     # Pad the waveform with 1 second of silence because sounddevice tends to cut them early
+    #     # when playing them. You shouldn't need to do that in your parsers.
+    #     wav = np.concatenate((wav, [0] * 16000))
+    #     print("\t%s" % text)
+    #     sd.play(wav, 16000, blocking=True)
+    # print("")
+    
+    return wavs, texts
+    
+    
+def process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str, 
+                      skip_existing: bool, hparams):
+    ## FOR REFERENCE:
+    # For you not to lose your head if you ever wish to change things here or implement your own
+    # synthesizer.
+    # - Both the audios and the mel spectrograms are saved as numpy arrays
+    # - There is no processing done to the audios that will be saved to disk beyond volume  
+    #   normalization (in split_on_silences)
+    # - However, pre-emphasis is applied to the audios before computing the mel spectrogram. This
+    #   is why we re-apply it on the audio on the side of the vocoder.
+    # - Librosa pads the waveform before computing the mel spectrogram. Here, the waveform is saved
+    #   without extra padding. This means that you won't have an exact relation between the length
+    #   of the wav and of the mel spectrogram. See the vocoder data loader.
+    
+    
+    # Skip existing utterances if needed
+    mel_fpath = out_dir.joinpath("mels", "mel-%s.npy" % basename)
+    wav_fpath = out_dir.joinpath("audio", "audio-%s.npy" % basename)
+    if skip_existing and mel_fpath.exists() and wav_fpath.exists():
+        return None
+
+    # Trim silence
+    if hparams.trim_silence:
+        wav = encoder.preprocess_wav(wav, normalize=False, trim_silence=True)
+    
+    # Skip utterances that are too short
+    if len(wav) < hparams.utterance_min_duration * hparams.sample_rate:
+        return None
+    
+    # Compute the mel spectrogram
+    mel_spectrogram = audio.melspectrogram(wav, hparams).astype(np.float32)
+    mel_frames = mel_spectrogram.shape[1]
+    
+    # Skip utterances that are too long
+    if mel_frames > hparams.max_mel_frames and hparams.clip_mels_length:
+        return None
+    
+    # Write the spectrogram, embed and audio to disk
+    np.save(mel_fpath, mel_spectrogram.T, allow_pickle=False)
+    np.save(wav_fpath, wav, allow_pickle=False)
+    
+    # Return a tuple describing this training example
+    return wav_fpath.name, mel_fpath.name, "embed-%s.npy" % basename, len(wav), mel_frames, text
+ 
+ 
+def embed_utterance(fpaths, encoder_model_fpath):
+    if not encoder.is_loaded():
+        encoder.load_model(encoder_model_fpath)
+
+    # Compute the speaker embedding of the utterance
+    wav_fpath, embed_fpath = fpaths
+    wav = np.load(wav_fpath)
+    wav = encoder.preprocess_wav(wav)
+    embed = encoder.embed_utterance(wav)
+    np.save(embed_fpath, embed, allow_pickle=False)
+    
+ 
+def create_embeddings(synthesizer_root: Path, encoder_model_fpath: Path, n_processes: int):
+    wav_dir = synthesizer_root.joinpath("audio")
+    metadata_fpath = synthesizer_root.joinpath("train.txt")
+    assert wav_dir.exists() and metadata_fpath.exists()
+    embed_dir = synthesizer_root.joinpath("embeds")
+    embed_dir.mkdir(exist_ok=True)
+    
+    # Gather the input wave filepath and the target output embed filepath
+    with metadata_fpath.open("r") as metadata_file:
+        metadata = [line.split("|") for line in metadata_file]
+        fpaths = [(wav_dir.joinpath(m[0]), embed_dir.joinpath(m[2])) for m in metadata]
+        
+    # TODO: improve on the multiprocessing, it's terrible. Disk I/O is the bottleneck here.
+    # Embed the utterances in separate threads
+    func = partial(embed_utterance, encoder_model_fpath=encoder_model_fpath)
+    job = Pool(n_processes).imap(func, fpaths)
+    list(tqdm(job, "Embedding", len(fpaths), unit="utterances"))
+    

synthesizer/synthesize.py

@@ -0,0 +1,97 @@
+import torch
+from torch.utils.data import DataLoader
+from synthesizer.hparams import hparams_debug_string
+from synthesizer.synthesizer_dataset import SynthesizerDataset, collate_synthesizer
+from synthesizer.models.tacotron import Tacotron
+from synthesizer.utils.text import text_to_sequence
+from synthesizer.utils.symbols import symbols
+import numpy as np
+from pathlib import Path
+from tqdm import tqdm
+import platform
+
+def run_synthesis(in_dir, out_dir, model_dir, hparams):
+    # This generates ground truth-aligned mels for vocoder training
+    synth_dir = Path(out_dir).joinpath("mels_gta")
+    synth_dir.mkdir(exist_ok=True)
+    print(hparams_debug_string())
+
+    # Check for GPU
+    if torch.cuda.is_available():
+        device = torch.device("cuda")
+        if hparams.synthesis_batch_size % torch.cuda.device_count() != 0:
+            raise ValueError("`hparams.synthesis_batch_size` must be evenly divisible by n_gpus!")
+    else:
+        device = torch.device("cpu")
+    print("Synthesizer using device:", device)
+
+    # Instantiate Tacotron model
+    model = Tacotron(embed_dims=hparams.tts_embed_dims,
+                     num_chars=len(symbols),
+                     encoder_dims=hparams.tts_encoder_dims,
+                     decoder_dims=hparams.tts_decoder_dims,
+                     n_mels=hparams.num_mels,
+                     fft_bins=hparams.num_mels,
+                     postnet_dims=hparams.tts_postnet_dims,
+                     encoder_K=hparams.tts_encoder_K,
+                     lstm_dims=hparams.tts_lstm_dims,
+                     postnet_K=hparams.tts_postnet_K,
+                     num_highways=hparams.tts_num_highways,
+                     dropout=0., # Use zero dropout for gta mels
+                     stop_threshold=hparams.tts_stop_threshold,
+                     speaker_embedding_size=hparams.speaker_embedding_size).to(device)
+
+    # Load the weights
+    model_dir = Path(model_dir)
+    model_fpath = model_dir.joinpath(model_dir.stem).with_suffix(".pt")
+    print("\nLoading weights at %s" % model_fpath)
+    model.load(model_fpath)
+    print("Tacotron weights loaded from step %d" % model.step)
+
+    # Synthesize using same reduction factor as the model is currently trained
+    r = np.int32(model.r)
+
+    # Set model to eval mode (disable gradient and zoneout)
+    model.eval()
+
+    # Initialize the dataset
+    in_dir = Path(in_dir)
+    metadata_fpath = in_dir.joinpath("train.txt")
+    mel_dir = in_dir.joinpath("mels")
+    embed_dir = in_dir.joinpath("embeds")
+
+    dataset = SynthesizerDataset(metadata_fpath, mel_dir, embed_dir, hparams)
+    data_loader = DataLoader(dataset,
+                             collate_fn=lambda batch: collate_synthesizer(batch, r, hparams),
+                             batch_size=hparams.synthesis_batch_size,
+                             num_workers=2 if platform.system() != "Windows" else 0,
+                             shuffle=False,
+                             pin_memory=True)
+
+    # Generate GTA mels
+    meta_out_fpath = Path(out_dir).joinpath("synthesized.txt")
+    with open(meta_out_fpath, "w") as file:
+        for i, (texts, mels, embeds, idx) in tqdm(enumerate(data_loader), total=len(data_loader)):
+            texts = texts.to(device)
+            mels = mels.to(device)
+            embeds = embeds.to(device)
+
+            # Parallelize model onto GPUS using workaround due to python bug
+            if device.type == "cuda" and torch.cuda.device_count() > 1:
+                _, mels_out, _ = data_parallel_workaround(model, texts, mels, embeds)
+            else:
+                _, mels_out, _, _ = model(texts, mels, embeds)
+
+            for j, k in enumerate(idx):
+                # Note: outputs mel-spectrogram files and target ones have same names, just different folders
+                mel_filename = Path(synth_dir).joinpath(dataset.metadata[k][1])
+                mel_out = mels_out[j].detach().cpu().numpy().T
+
+                # Use the length of the ground truth mel to remove padding from the generated mels
+                mel_out = mel_out[:int(dataset.metadata[k][4])]
+
+                # Write the spectrogram to disk
+                np.save(mel_filename, mel_out, allow_pickle=False)
+
+                # Write metadata into the synthesized file
+                file.write("|".join(dataset.metadata[k]))

synthesizer/synthesizer_dataset.py

@@ -0,0 +1,92 @@
+import torch
+from torch.utils.data import Dataset
+import numpy as np
+from pathlib import Path
+from synthesizer.utils.text import text_to_sequence
+
+
+class SynthesizerDataset(Dataset):
+    def __init__(self, metadata_fpath: Path, mel_dir: Path, embed_dir: Path, hparams):
+        print("Using inputs from:\n\t%s\n\t%s\n\t%s" % (metadata_fpath, mel_dir, embed_dir))
+        
+        with metadata_fpath.open("r") as metadata_file:
+            metadata = [line.split("|") for line in metadata_file]
+        
+        mel_fnames = [x[1] for x in metadata if int(x[4])]
+        mel_fpaths = [mel_dir.joinpath(fname) for fname in mel_fnames]
+        embed_fnames = [x[2] for x in metadata if int(x[4])]
+        embed_fpaths = [embed_dir.joinpath(fname) for fname in embed_fnames]
+        self.samples_fpaths = list(zip(mel_fpaths, embed_fpaths))
+        self.samples_texts = [x[5].strip() for x in metadata if int(x[4])]
+        self.metadata = metadata
+        self.hparams = hparams
+        
+        print("Found %d samples" % len(self.samples_fpaths))
+    
+    def __getitem__(self, index):  
+        # Sometimes index may be a list of 2 (not sure why this happens)
+        # If that is the case, return a single item corresponding to first element in index
+        if index is list:
+            index = index[0]
+
+        mel_path, embed_path = self.samples_fpaths[index]
+        mel = np.load(mel_path).T.astype(np.float32)
+        
+        # Load the embed
+        embed = np.load(embed_path)
+
+        # Get the text and clean it
+        text = text_to_sequence(self.samples_texts[index], self.hparams.tts_cleaner_names)
+        
+        # Convert the list returned by text_to_sequence to a numpy array
+        text = np.asarray(text).astype(np.int32)
+
+        return text, mel.astype(np.float32), embed.astype(np.float32), index
+
+    def __len__(self):
+        return len(self.samples_fpaths)
+
+
+def collate_synthesizer(batch, r, hparams):
+    # Text
+    x_lens = [len(x[0]) for x in batch]
+    max_x_len = max(x_lens)
+
+    chars = [pad1d(x[0], max_x_len) for x in batch]
+    chars = np.stack(chars)
+
+    # Mel spectrogram
+    spec_lens = [x[1].shape[-1] for x in batch]
+    max_spec_len = max(spec_lens) + 1 
+    if max_spec_len % r != 0:
+        max_spec_len += r - max_spec_len % r 
+
+    # WaveRNN mel spectrograms are normalized to [0, 1] so zero padding adds silence
+    # By default, SV2TTS uses symmetric mels, where -1*max_abs_value is silence.
+    if hparams.symmetric_mels:
+        mel_pad_value = -1 * hparams.max_abs_value
+    else:
+        mel_pad_value = 0
+
+    mel = [pad2d(x[1], max_spec_len, pad_value=mel_pad_value) for x in batch]
+    mel = np.stack(mel)
+
+    # Speaker embedding (SV2TTS)
+    embeds = [x[2] for x in batch]
+
+    # Index (for vocoder preprocessing)
+    indices = [x[3] for x in batch]
+
+
+    # Convert all to tensor
+    chars = torch.tensor(chars).long()
+    mel = torch.tensor(mel)
+    embeds = torch.tensor(embeds)
+
+    return chars, mel, embeds, indices
+
+def pad1d(x, max_len, pad_value=0):
+    return np.pad(x, (0, max_len - len(x)), mode="constant", constant_values=pad_value)
+
+def pad2d(x, max_len, pad_value=0):
+    return np.pad(x, ((0, 0), (0, max_len - x.shape[-1])), mode="constant", constant_values=pad_value)

synthesizer/train.py

@@ -0,0 +1,269 @@
+import torch
+import torch.nn.functional as F
+from torch import optim
+from torch.utils.data import DataLoader
+from synthesizer import audio
+from synthesizer.models.tacotron import Tacotron
+from synthesizer.synthesizer_dataset import SynthesizerDataset, collate_synthesizer
+from synthesizer.utils import ValueWindow, data_parallel_workaround
+from synthesizer.utils.plot import plot_spectrogram
+from synthesizer.utils.symbols import symbols
+from synthesizer.utils.text import sequence_to_text
+from vocoder.display import *
+from datetime import datetime
+import numpy as np
+from pathlib import Path
+import sys
+import time
+import platform
+
+
+def np_now(x: torch.Tensor): return x.detach().cpu().numpy()
+
+def time_string():
+    return datetime.now().strftime("%Y-%m-%d %H:%M")
+
+def train(run_id: str, syn_dir: str, models_dir: str, save_every: int,
+         backup_every: int, force_restart:bool, hparams):
+
+    syn_dir = Path(syn_dir)
+    models_dir = Path(models_dir)
+    models_dir.mkdir(exist_ok=True)
+
+    model_dir = models_dir.joinpath(run_id)
+    plot_dir = model_dir.joinpath("plots")
+    wav_dir = model_dir.joinpath("wavs")
+    mel_output_dir = model_dir.joinpath("mel-spectrograms")
+    meta_folder = model_dir.joinpath("metas")
+    model_dir.mkdir(exist_ok=True)
+    plot_dir.mkdir(exist_ok=True)
+    wav_dir.mkdir(exist_ok=True)
+    mel_output_dir.mkdir(exist_ok=True)
+    meta_folder.mkdir(exist_ok=True)
+    
+    weights_fpath = model_dir.joinpath(run_id).with_suffix(".pt")
+    metadata_fpath = syn_dir.joinpath("train.txt")
+    
+    print("Checkpoint path: {}".format(weights_fpath))
+    print("Loading training data from: {}".format(metadata_fpath))
+    print("Using model: Tacotron")
+    
+    # Book keeping
+    step = 0
+    time_window = ValueWindow(100)
+    loss_window = ValueWindow(100)
+    
+    
+    # From WaveRNN/train_tacotron.py
+    if torch.cuda.is_available():
+        device = torch.device("cuda")
+
+        for session in hparams.tts_schedule:
+            _, _, _, batch_size = session
+            if batch_size % torch.cuda.device_count() != 0:
+                raise ValueError("`batch_size` must be evenly divisible by n_gpus!")
+    else:
+        device = torch.device("cpu")
+    print("Using device:", device)
+
+    # Instantiate Tacotron Model
+    print("\nInitialising Tacotron Model...\n")
+    model = Tacotron(embed_dims=hparams.tts_embed_dims,
+                     num_chars=len(symbols),
+                     encoder_dims=hparams.tts_encoder_dims,
+                     decoder_dims=hparams.tts_decoder_dims,
+                     n_mels=hparams.num_mels,
+                     fft_bins=hparams.num_mels,
+                     postnet_dims=hparams.tts_postnet_dims,
+                     encoder_K=hparams.tts_encoder_K,
+                     lstm_dims=hparams.tts_lstm_dims,
+                     postnet_K=hparams.tts_postnet_K,
+                     num_highways=hparams.tts_num_highways,
+                     dropout=hparams.tts_dropout,
+                     stop_threshold=hparams.tts_stop_threshold,
+                     speaker_embedding_size=hparams.speaker_embedding_size).to(device)
+
+    # Initialize the optimizer
+    optimizer = optim.Adam(model.parameters())
+
+    # Load the weights
+    if force_restart or not weights_fpath.exists():
+        print("\nStarting the training of Tacotron from scratch\n")
+        model.save(weights_fpath)
+
+        # Embeddings metadata
+        char_embedding_fpath = meta_folder.joinpath("CharacterEmbeddings.tsv")
+        with open(char_embedding_fpath, "w", encoding="utf-8") as f:
+            for symbol in symbols:
+                if symbol == " ":
+                    symbol = "\\s"  # For visual purposes, swap space with \s
+
+                f.write("{}\n".format(symbol))
+
+    else:
+        print("\nLoading weights at %s" % weights_fpath)
+        model.load(weights_fpath, optimizer)
+        print("Tacotron weights loaded from step %d" % model.step)
+    
+    # Initialize the dataset
+    metadata_fpath = syn_dir.joinpath("train.txt")
+    mel_dir = syn_dir.joinpath("mels")
+    embed_dir = syn_dir.joinpath("embeds")
+    dataset = SynthesizerDataset(metadata_fpath, mel_dir, embed_dir, hparams)
+    test_loader = DataLoader(dataset,
+                             batch_size=1,
+                             shuffle=True,
+                             pin_memory=True)
+
+    for i, session in enumerate(hparams.tts_schedule):
+        current_step = model.get_step()
+
+        r, lr, max_step, batch_size = session
+
+        training_steps = max_step - current_step
+
+        # Do we need to change to the next session?
+        if current_step >= max_step:
+            # Are there no further sessions than the current one?
+            if i == len(hparams.tts_schedule) - 1:
+                # We have completed training. Save the model and exit
+                model.save(weights_fpath, optimizer)
+                break
+            else:
+                # There is a following session, go to it
+                continue
+
+        model.r = r
+
+        # Begin the training
+        simple_table([(f"Steps with r={r}", str(training_steps // 1000) + "k Steps"),
+                      ("Batch Size", batch_size),
+                      ("Learning Rate", lr),
+                      ("Outputs/Step (r)", model.r)])
+
+        for p in optimizer.param_groups:
+            p["lr"] = lr
+
+        data_loader = DataLoader(dataset,
+                                 collate_fn=lambda batch: collate_synthesizer(batch, r, hparams),
+                                 batch_size=batch_size,
+                                 num_workers=2 if platform.system() != "Windows" else 0,
+                                 shuffle=True,
+                                 pin_memory=True)
+
+        total_iters = len(dataset) 
+        steps_per_epoch = np.ceil(total_iters / batch_size).astype(np.int32)
+        epochs = np.ceil(training_steps / steps_per_epoch).astype(np.int32)
+
+        for epoch in range(1, epochs+1):
+            for i, (texts, mels, embeds, idx) in enumerate(data_loader, 1):
+                start_time = time.time()
+
+                # Generate stop tokens for training
+                stop = torch.ones(mels.shape[0], mels.shape[2])
+                for j, k in enumerate(idx):
+                    stop[j, :int(dataset.metadata[k][4])-1] = 0
+
+                texts = texts.to(device)
+                mels = mels.to(device)
+                embeds = embeds.to(device)
+                stop = stop.to(device)
+
+                # Forward pass
+                # Parallelize model onto GPUS using workaround due to python bug
+                if device.type == "cuda" and torch.cuda.device_count() > 1:
+                    m1_hat, m2_hat, attention, stop_pred = data_parallel_workaround(model, texts,
+                                                                                    mels, embeds)
+                else:
+                    m1_hat, m2_hat, attention, stop_pred = model(texts, mels, embeds)
+
+                # Backward pass
+                m1_loss = F.mse_loss(m1_hat, mels) + F.l1_loss(m1_hat, mels)
+                m2_loss = F.mse_loss(m2_hat, mels)
+                stop_loss = F.binary_cross_entropy(stop_pred, stop)
+
+                loss = m1_loss + m2_loss + stop_loss
+
+                optimizer.zero_grad()
+                loss.backward()
+
+                if hparams.tts_clip_grad_norm is not None:
+                    grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), hparams.tts_clip_grad_norm)
+                    if np.isnan(grad_norm.cpu()):
+                        print("grad_norm was NaN!")
+
+                optimizer.step()
+
+                time_window.append(time.time() - start_time)
+                loss_window.append(loss.item())
+
+                step = model.get_step()
+                k = step // 1000
+
+                msg = f"| Epoch: {epoch}/{epochs} ({i}/{steps_per_epoch}) | Loss: {loss_window.average:#.4} | {1./time_window.average:#.2} steps/s | Step: {k}k | "
+                stream(msg)
+
+                # Backup or save model as appropriate
+                if backup_every != 0 and step % backup_every == 0 : 
+                    backup_fpath = Path("{}/{}_{}k.pt".format(str(weights_fpath.parent), run_id, k))
+                    model.save(backup_fpath, optimizer)
+
+                if save_every != 0 and step % save_every == 0 : 
+                    # Must save latest optimizer state to ensure that resuming training
+                    # doesn't produce artifacts
+                    model.save(weights_fpath, optimizer)
+
+                # Evaluate model to generate samples
+                epoch_eval = hparams.tts_eval_interval == -1 and i == steps_per_epoch  # If epoch is done
+                step_eval = hparams.tts_eval_interval > 0 and step % hparams.tts_eval_interval == 0  # Every N steps
+                if epoch_eval or step_eval:
+                    for sample_idx in range(hparams.tts_eval_num_samples):
+                        # At most, generate samples equal to number in the batch
+                        if sample_idx + 1 <= len(texts):
+                            # Remove padding from mels using frame length in metadata
+                            mel_length = int(dataset.metadata[idx[sample_idx]][4])
+                            mel_prediction = np_now(m2_hat[sample_idx]).T[:mel_length]
+                            target_spectrogram = np_now(mels[sample_idx]).T[:mel_length]
+                            attention_len = mel_length // model.r
+
+                            eval_model(attention=np_now(attention[sample_idx][:, :attention_len]),
+                                       mel_prediction=mel_prediction,
+                                       target_spectrogram=target_spectrogram,
+                                       input_seq=np_now(texts[sample_idx]),
+                                       step=step,
+                                       plot_dir=plot_dir,
+                                       mel_output_dir=mel_output_dir,
+                                       wav_dir=wav_dir,
+                                       sample_num=sample_idx + 1,
+                                       loss=loss,
+                                       hparams=hparams)
+
+                # Break out of loop to update training schedule
+                if step >= max_step:
+                    break
+
+            # Add line break after every epoch
+            print("")
+
+def eval_model(attention, mel_prediction, target_spectrogram, input_seq, step,
+               plot_dir, mel_output_dir, wav_dir, sample_num, loss, hparams):
+    # Save some results for evaluation
+    attention_path = str(plot_dir.joinpath("attention_step_{}_sample_{}".format(step, sample_num)))
+    save_attention(attention, attention_path)
+
+    # save predicted mel spectrogram to disk (debug)
+    mel_output_fpath = mel_output_dir.joinpath("mel-prediction-step-{}_sample_{}.npy".format(step, sample_num))
+    np.save(str(mel_output_fpath), mel_prediction, allow_pickle=False)
+
+    # save griffin lim inverted wav for debug (mel -> wav)
+    wav = audio.inv_mel_spectrogram(mel_prediction.T, hparams)
+    wav_fpath = wav_dir.joinpath("step-{}-wave-from-mel_sample_{}.wav".format(step, sample_num))
+    audio.save_wav(wav, str(wav_fpath), sr=hparams.sample_rate)
+
+    # save real and predicted mel-spectrogram plot to disk (control purposes)
+    spec_fpath = plot_dir.joinpath("step-{}-mel-spectrogram_sample_{}.png".format(step, sample_num))
+    title_str = "{}, {}, step={}, loss={:.5f}".format("Tacotron", time_string(), step, loss)
+    plot_spectrogram(mel_prediction, str(spec_fpath), title=title_str,
+                     target_spectrogram=target_spectrogram,
+                     max_len=target_spectrogram.size // hparams.num_mels)
+    print("Input at step {}: {}".format(step, sequence_to_text(input_seq)))

synthesizer/utils/__init__.py

@@ -0,0 +1,45 @@
+import torch
+
+
+_output_ref = None
+_replicas_ref = None
+
+def data_parallel_workaround(model, *input):
+    global _output_ref
+    global _replicas_ref
+    device_ids = list(range(torch.cuda.device_count()))
+    output_device = device_ids[0]
+    replicas = torch.nn.parallel.replicate(model, device_ids)
+    # input.shape = (num_args, batch, ...)
+    inputs = torch.nn.parallel.scatter(input, device_ids)
+    # inputs.shape = (num_gpus, num_args, batch/num_gpus, ...)
+    replicas = replicas[:len(inputs)]
+    outputs = torch.nn.parallel.parallel_apply(replicas, inputs)
+    y_hat = torch.nn.parallel.gather(outputs, output_device)
+    _output_ref = outputs
+    _replicas_ref = replicas
+    return y_hat
+
+
+class ValueWindow():
+  def __init__(self, window_size=100):
+    self._window_size = window_size
+    self._values = []
+
+  def append(self, x):
+    self._values = self._values[-(self._window_size - 1):] + [x]
+
+  @property
+  def sum(self):
+    return sum(self._values)
+
+  @property
+  def count(self):
+    return len(self._values)
+
+  @property
+  def average(self):
+    return self.sum / max(1, self.count)
+
+  def reset(self):
+    self._values = []

synthesizer/utils/_cmudict.py

@@ -0,0 +1,62 @@
+import re
+
+valid_symbols = [
+  "AA", "AA0", "AA1", "AA2", "AE", "AE0", "AE1", "AE2", "AH", "AH0", "AH1", "AH2",
+  "AO", "AO0", "AO1", "AO2", "AW", "AW0", "AW1", "AW2", "AY", "AY0", "AY1", "AY2",
+  "B", "CH", "D", "DH", "EH", "EH0", "EH1", "EH2", "ER", "ER0", "ER1", "ER2", "EY",
+  "EY0", "EY1", "EY2", "F", "G", "HH", "IH", "IH0", "IH1", "IH2", "IY", "IY0", "IY1",
+  "IY2", "JH", "K", "L", "M", "N", "NG", "OW", "OW0", "OW1", "OW2", "OY", "OY0",
+  "OY1", "OY2", "P", "R", "S", "SH", "T", "TH", "UH", "UH0", "UH1", "UH2", "UW",
+  "UW0", "UW1", "UW2", "V", "W", "Y", "Z", "ZH"
+]
+
+_valid_symbol_set = set(valid_symbols)
+
+
+class CMUDict:
+  """Thin wrapper around CMUDict data. http://www.speech.cs.cmu.edu/cgi-bin/cmudict"""
+  def __init__(self, file_or_path, keep_ambiguous=True):
+    if isinstance(file_or_path, str):
+      with open(file_or_path, encoding="latin-1") as f:
+        entries = _parse_cmudict(f)
+    else:
+      entries = _parse_cmudict(file_or_path)
+    if not keep_ambiguous:
+      entries = {word: pron for word, pron in entries.items() if len(pron) == 1}
+    self._entries = entries
+
+
+  def __len__(self):
+    return len(self._entries)
+
+
+  def lookup(self, word):
+    """Returns list of ARPAbet pronunciations of the given word."""
+    return self._entries.get(word.upper())
+
+
+
+_alt_re = re.compile(r"\([0-9]+\)")
+
+
+def _parse_cmudict(file):
+  cmudict = {}
+  for line in file:
+    if len(line) and (line[0] >= "A" and line[0] <= "Z" or line[0] == "'"):
+      parts = line.split("  ")
+      word = re.sub(_alt_re, "", parts[0])
+      pronunciation = _get_pronunciation(parts[1])
+      if pronunciation:
+        if word in cmudict:
+          cmudict[word].append(pronunciation)
+        else:
+          cmudict[word] = [pronunciation]
+  return cmudict
+
+
+def _get_pronunciation(s):
+  parts = s.strip().split(" ")
+  for part in parts:
+    if part not in _valid_symbol_set:
+      return None
+  return " ".join(parts)

synthesizer/utils/cleaners.py

@@ -0,0 +1,88 @@
+"""
+Cleaners are transformations that run over the input text at both training and eval time.
+
+Cleaners can be selected by passing a comma-delimited list of cleaner names as the "cleaners"
+hyperparameter. Some cleaners are English-specific. You"ll typically want to use:
+  1. "english_cleaners" for English text
+  2. "transliteration_cleaners" for non-English text that can be transliterated to ASCII using
+     the Unidecode library (https://pypi.python.org/pypi/Unidecode)
+  3. "basic_cleaners" if you do not want to transliterate (in this case, you should also update
+     the symbols in symbols.py to match your data).
+"""
+
+import re
+from unidecode import unidecode
+from .numbers import normalize_numbers
+
+# Regular expression matching whitespace:
+_whitespace_re = re.compile(r"\s+")
+
+# List of (regular expression, replacement) pairs for abbreviations:
+_abbreviations = [(re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1]) for x in [
+  ("mrs", "misess"),
+  ("mr", "mister"),
+  ("dr", "doctor"),
+  ("st", "saint"),
+  ("co", "company"),
+  ("jr", "junior"),
+  ("maj", "major"),
+  ("gen", "general"),
+  ("drs", "doctors"),
+  ("rev", "reverend"),
+  ("lt", "lieutenant"),
+  ("hon", "honorable"),
+  ("sgt", "sergeant"),
+  ("capt", "captain"),
+  ("esq", "esquire"),
+  ("ltd", "limited"),
+  ("col", "colonel"),
+  ("ft", "fort"),
+]]
+
+
+def expand_abbreviations(text):
+  for regex, replacement in _abbreviations:
+    text = re.sub(regex, replacement, text)
+  return text
+
+
+def expand_numbers(text):
+  return normalize_numbers(text)
+
+
+def lowercase(text):
+  """lowercase input tokens."""
+  return text.lower()
+
+
+def collapse_whitespace(text):
+  return re.sub(_whitespace_re, " ", text)
+
+
+def convert_to_ascii(text):
+  return unidecode(text)
+
+
+def basic_cleaners(text):
+  """Basic pipeline that lowercases and collapses whitespace without transliteration."""
+  text = lowercase(text)
+  text = collapse_whitespace(text)
+  return text
+
+
+def transliteration_cleaners(text):
+  """Pipeline for non-English text that transliterates to ASCII."""
+  text = convert_to_ascii(text)
+  text = lowercase(text)
+  text = collapse_whitespace(text)
+  return text
+
+
+def english_cleaners(text):
+  """Pipeline for English text, including number and abbreviation expansion."""
+  text = convert_to_ascii(text)
+  text = lowercase(text)
+  text = expand_numbers(text)
+  text = expand_abbreviations(text)
+  text = collapse_whitespace(text)
+  return text

synthesizer/utils/numbers.py

@@ -0,0 +1,68 @@
+import re
+import inflect
+
+_inflect = inflect.engine()
+_comma_number_re = re.compile(r"([0-9][0-9\,]+[0-9])")
+_decimal_number_re = re.compile(r"([0-9]+\.[0-9]+)")
+_pounds_re = re.compile(r"£([0-9\,]*[0-9]+)")
+_dollars_re = re.compile(r"\$([0-9\.\,]*[0-9]+)")
+_ordinal_re = re.compile(r"[0-9]+(st|nd|rd|th)")
+_number_re = re.compile(r"[0-9]+")
+
+
+def _remove_commas(m):
+  return m.group(1).replace(",", "")
+
+
+def _expand_decimal_point(m):
+  return m.group(1).replace(".", " point ")
+
+
+def _expand_dollars(m):
+  match = m.group(1)
+  parts = match.split(".")
+  if len(parts) > 2:
+    return match + " dollars"  # Unexpected format
+  dollars = int(parts[0]) if parts[0] else 0
+  cents = int(parts[1]) if len(parts) > 1 and parts[1] else 0
+  if dollars and cents:
+    dollar_unit = "dollar" if dollars == 1 else "dollars"
+    cent_unit = "cent" if cents == 1 else "cents"
+    return "%s %s, %s %s" % (dollars, dollar_unit, cents, cent_unit)
+  elif dollars:
+    dollar_unit = "dollar" if dollars == 1 else "dollars"
+    return "%s %s" % (dollars, dollar_unit)
+  elif cents:
+    cent_unit = "cent" if cents == 1 else "cents"
+    return "%s %s" % (cents, cent_unit)
+  else:
+    return "zero dollars"
+
+
+def _expand_ordinal(m):
+  return _inflect.number_to_words(m.group(0))
+
+
+def _expand_number(m):
+  num = int(m.group(0))
+  if num > 1000 and num < 3000:
+    if num == 2000:
+      return "two thousand"
+    elif num > 2000 and num < 2010:
+      return "two thousand " + _inflect.number_to_words(num % 100)
+    elif num % 100 == 0:
+      return _inflect.number_to_words(num // 100) + " hundred"
+    else:
+      return _inflect.number_to_words(num, andword="", zero="oh", group=2).replace(", ", " ")
+  else:
+    return _inflect.number_to_words(num, andword="")
+
+
+def normalize_numbers(text):
+  text = re.sub(_comma_number_re, _remove_commas, text)
+  text = re.sub(_pounds_re, r"\1 pounds", text)
+  text = re.sub(_dollars_re, _expand_dollars, text)
+  text = re.sub(_decimal_number_re, _expand_decimal_point, text)
+  text = re.sub(_ordinal_re, _expand_ordinal, text)
+  text = re.sub(_number_re, _expand_number, text)
+  return text