Create app.py

app.py

@@ -0,0 +1,181 @@
+import torch
+from torch.utils.data import DataLoader
+import numpy as np
+from tqdm import tqdm
+from transformers import SpeechT5HifiGan
+from datasets import load_dataset
+from tqdm import tqdm
+import soundfile as sf
+import librosa
+
+dataset = load_dataset('pourmand1376/asr-farsi-youtube-chunked-10-seconds', split = "test")
+
+
+import librosa
+from datasets import load_dataset, Audio
+
+def resample_audio(example):
+    # Resample to 16 kHz
+    y_resampled = librosa.resample(example["audio"]["array"], orig_sr=example["audio"]["sampling_rate"], target_sr=16000)
+    
+    # Update the example with the resampled audio and new sample rate
+    example["audio"]["array"] = y_resampled
+    example["audio"]["sampling_rate"] = 16000
+    
+    return example
+
+
+dataset = dataset.select(range(1000))
+dataset = dataset.map(resample_audio)
+
+
+import torch
+from torch.utils.data import DataLoader
+import numpy as np
+from tqdm import tqdm
+from transformers import SpeechT5HifiGan
+from datasets import load_dataset
+from tqdm import tqdm
+import soundfile as sf
+import librosa
+
+
+
+def set_seed(seed):
+  torch.manual_seed(seed)
+  if torch.cuda.is_available():
+    torch.cuda.manual_seed_all(seed)
+
+set_seed(1)
+# Load model directly
+from transformers import AutoProcessor, AutoModelForTextToSpectrogram
+
+processor = AutoProcessor.from_pretrained("Alidr79/speecht5_v3_youtube")
+model = AutoModelForTextToSpectrogram.from_pretrained("Alidr79/speecht5_v3_youtube")
+
+
+from speechbrain.inference.classifiers import EncoderClassifier
+import os
+
+spk_model_name = "speechbrain/spkrec-xvect-voxceleb"
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+speaker_model = EncoderClassifier.from_hparams(
+    source=spk_model_name,
+    run_opts={"device": device},
+    savedir=os.path.join("/tmp", spk_model_name),
+)
+
+
+def create_speaker_embedding(waveform):
+    with torch.no_grad():
+        speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform))
+        speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
+        speaker_embeddings = speaker_embeddings.squeeze().cpu().numpy()
+    return speaker_embeddings
+
+vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+
+from PersianG2p import Persian_g2p_converter
+from scipy.io import wavfile
+import soundfile as sf
+
+
+PersianG2Pconverter = Persian_g2p_converter(use_large = True)
+
+import noisereduce as nr
+
+def denoise_audio(audio, sr):
+    # Perform noise reduction
+    denoised_audio = nr.reduce_noise(y=audio, sr=sr)
+    return denoised_audio
+
+
+import noisereduce as nr
+from pydub import AudioSegment
+def match_target_amplitude(sound, target_dBFS):
+    change_in_dBFS = target_dBFS - sound.dBFS
+    return sound.apply_gain(change_in_dBFS)
+
+import librosa
+def tts_fn(slider_value, input_text):
+    audio_embedding = dataset[slider_value]['audio']['array']
+    sample_rate_embedding = dataset[slider_value]['audio']['sampling_rate']
+    if sample_rate_embedding != 16000:
+        audio_embedding = librosa.resample(audio_embedding, orig_sr=sample_rate_embedding, target_sr=16_000)
+        
+    
+    with torch.no_grad():
+        speaker_embedding = create_speaker_embedding(audio_embedding)
+        speaker_embedding = torch.tensor(speaker_embedding).unsqueeze(0)
+
+    phonemes = PersianG2Pconverter.transliterate(input_text, tidy = False, secret = True)
+    # text = "</s>"
+    # for i in phonemes.replace(' .', '').split(" "):
+    #     text += i + " <pad> "
+
+    text = phonemes
+    
+    print("sentence:", input_text)
+    print("sentence phonemes:", text)
+
+    with torch.no_grad():
+        inputs = processor(text = text, return_tensors="pt")
+
+    with torch.no_grad():
+        spectrogram = model.generate_speech(inputs["input_ids"], speaker_embedding, minlenratio = 2, maxlenratio = 4, threshold = 0.35)
+
+    with torch.no_grad():
+        speech = vocoder(spectrogram)
+
+    speech = speech.numpy().reshape(-1)
+    speech_denoised = denoise_audio(speech, 16000)
+    sf.write("in_speech.wav", speech_denoised, 16000)
+
+    sound = AudioSegment.from_wav("in_speech.wav", "wav")
+    normalized_sound = match_target_amplitude(sound, -20.0)
+    normalized_sound.export("out_sound.wav", format="wav")
+
+    sample_rate_out, audio_out = wavfile.read("out_sound.wav")
+
+    assert sample_rate_out == 16_000
+
+    return 16000, (audio_out.reshape(-1)).astype(np.int16)
+
+
+def master_fn(slider_value, input_text):
+    if "." not in input_text:
+        input_text += '.'
+        
+    all_speech = []
+    for sentence in input_text.split("."):
+        sampling_rate_response, audio_chunk_response = tts_fn(slider_value, sentence)
+        all_speech.append(audio_chunk_response)
+       
+    audio_response = np.concatenate(all_speech)
+    return sampling_rate_response, audio_response
+
+import gradio as gr
+
+slider = gr.Slider(
+    minimum=0,
+    maximum=(len(dataset)-1),
+    value=600,
+    step=1,
+    label="Select a speaker(Good examples : 600, 604, 910, 7, 13)"
+)
+
+# Create the text input component
+text_input = gr.Textbox(
+    label="Enter some text",
+    placeholder="Type something here..."
+)
+
+
+demo = gr.Interface(
+    fn = master_fn,
+    inputs=[slider, text_input],  # List of inputs
+    outputs = "audio"
+)
+
+demo.launch()