""" TODO: + [x] Load Configuration + [ ] Checking + [ ] Better saving directory """ import numpy as np from pathlib import Path import torch.nn as nn import torch import torchaudio from transformers import pipeline from pathlib import Path import pdb # local import import sys from espnet2.bin.tts_inference import Text2Speech from transformers import AutoTokenizer, AutoFeatureExtractor, AutoModelForCTC device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") sys.path.append("src") import gradio as gr from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq processor = AutoProcessor.from_pretrained("KevinGeng/whipser_medium_en_PAL300_step25") model = AutoModelForSpeechSeq2Seq.from_pretrained("KevinGeng/whipser_medium_en_PAL300_step25") transcriber = pipeline("automatic-speech-recognition", model="KevinGeng/whipser_medium_en_PAL300_step25") # Text2Mel models # @title English multi-speaker pretrained model { run: "auto" } lang = "English" vits_tag = "kan-bayashi/libritts_xvector_vits" ft2_tag = "kan-bayashi/libritts_xvector_conformer_fastspeech2" transformer_tag = "kan-bayashi/libritts_xvector_transformer" # !!! vits needs no vocoder !!! # Local Text2Mel models vits_config_local = "TTS_models/libritts_xvector_vits/config.yaml" vits_model_local = "TTS_models/libritts_xvector_vits/train.total_count.ave_10best.pth" # TODO ft2_config_local = "" ft2_model_local= "" transformer_config_local = "" transformer_config_local = "" # Vocoders vocoder_tag = "parallel_wavegan/vctk_parallel_wavegan.v1.long" # @param ["none", "parallel_wavegan/vctk_parallel_wavegan.v1.long", "parallel_wavegan/vctk_multi_band_melgan.v2", "parallel_wavegan/vctk_style_melgan.v1", "parallel_wavegan/vctk_hifigan.v1", "parallel_wavegan/libritts_parallel_wavegan.v1.long", "parallel_wavegan/libritts_multi_band_melgan.v2", "parallel_wavegan/libritts_hifigan.v1", "parallel_wavegan/libritts_style_melgan.v1"] {type:"string"} hifigan_vocoder_tag = "parallel_wavegan/parallel_wavegan/libritts_hifigan.v1" # @param ["none", "parallel_wavegan/vctk_parallel_wavegan.v1.long", "parallel_wavegan/vctk_multi_band_melgan.v2", "parallel_wavegan/vctk_style_melgan.v1", "parallel_wavegan/vctk_hifigan.v1", "parallel_wavegan/libritts_parallel_wavegan.v1.long", "parallel_wavegan/libritts_multi_band_melgan.v2", "parallel_wavegan/libritts_hifigan.v1", "parallel_wavegan/libritts_style_melgan.v1"] {type:"string"} # Local Vocoders ## Make sure the use parallel_wavegan as prefix (PWG feature) vocoder_tag_local = "parallel_wavegan/vctk_parallel_wavegan.v1.long" hifigan_vocoder_tag_local = "parallel_wavegan/libritts_hifigan.v1" from espnet2.bin.tts_inference import Text2Speech from espnet2.utils.types import str_or_none # local import text2speech = Text2Speech.from_pretrained( train_config = vits_config_local, model_file=vits_model_local, device="cuda", use_att_constraint=False, backward_window=1, forward_window=3, speed_control_alpha=1.0, ) # # Fastspeech2 # ft2_text2speech = Text2Speech.from_pretrained( # model_tag=ft2_tag, # vocoder_tag=str_or_none(vocoder_tag_local), # device="cuda", # use_att_constraint=False, # backward_window=1, # forward_window=3, # speed_control_alpha=1.0, # ) # # Fastspeech2 + hifigan # ft2_text2speech_hifi = Text2Speech.from_pretrained( # model_tag=ft2_tag, # vocoder_tag=str_or_none(hifigan_vocoder_tag_local), # device="cuda", # use_att_constraint=False, # backward_window=1, # forward_window=3, # speed_control_alpha=1.0, # ) # # transformer tag # transformer_text2speech = Text2Speech.from_pretrained( # model_tag=transformer_tag, # vocoder_tag=str_or_none(vocoder_tag_local), # device="cuda", # use_att_constraint=False, # backward_window=1, # forward_window=3, # speed_control_alpha=1.0, # ) import glob import os import numpy as np import kaldiio # Get model directory path # from espnet_model_zoo.downloader import ModelDownloader # d = ModelDownloader() # model_dir = os.path.dirname(d.download_and_unpack(tag)["train_config"]) # Speaker x-vector selection xvector_ark = [ p for p in glob.glob( f"xvector/test-clean/spk_xvector.ark", recursive=True ) if "test" in p ][0] xvectors = {k: v for k, v in kaldiio.load_ark(xvector_ark)} spks = list(xvectors.keys()) male_spks = { "Male1": "260_123286", "Male2": "1320_122612", "Male3": "672_122797" } female_spks = {"Female1": "5683_32865", "Female2": "121_121726", "Female3": "8463_287645"} spks = dict(male_spks, **female_spks) spk_names = sorted(spks.keys()) def ASRTTS(audio_file, spk_name, ref_text=""): spk = spks[spk_name] spembs = xvectors[spk] if ref_text == "": reg_text = transcriber(audio_file)["text"] else: reg_text = ref_text speech, sr = torchaudio.load( audio_file, channels_first=True ) # Mono channel wav_tensor_spembs = text2speech( text=reg_text, speech=speech, spembs=spembs )["wav"] wav_numpy = wav_tensor_spembs.unsqueeze(1).to("cpu") sample_rate = 22050 save_id = ( "./wav/" + Path(audio_file).stem + "_" + spk_name + "_spkembs.wav" ) torchaudio.save( save_id, src=wav_tensor_spembs.unsqueeze(0).to("cpu"), sample_rate=22050, ) return save_id, reg_text def ASRTTS_clean(audio_file, spk_name): spk = spks[spk_name] spembs = xvectors[spk] reg_text = transcriber(audio_file)["text"] speech, sr = torchaudio.load( audio_file, channels_first=True ) # Mono channel wav_tensor_spembs = text2speech( text=reg_text, speech=speech, spembs=spembs )["wav"] wav_numpy = wav_tensor_spembs.unsqueeze(1).to("cpu") sample_rate = 22050 save_id = ( "./wav/" + Path(audio_file).stem + "_" + spk_name + "_spkembs.wav" ) torchaudio.save( save_id, src=wav_tensor_spembs.unsqueeze(0).to("cpu"), sample_rate=22050, ) return save_id def ft2_ASRTTS_clean(audio_file, spk_name): spk = spks[spk_name] spembs = xvectors[spk] reg_text = transcriber(audio_file)["text"] speech, sr = torchaudio.load( audio_file, channels_first=True ) # Mono channel wav_tensor_spembs = ft2_text2speech( text=reg_text, speech=speech, spembs=spembs )["wav"] wav_numpy = wav_tensor_spembs.unsqueeze(1).to("cpu") sample_rate = 22050 save_id = ( "./wav/" + Path(audio_file).stem + "_fs2_" + spk_name + "_spkembs.wav" ) torchaudio.save( save_id, src=wav_tensor_spembs.unsqueeze(0).to("cpu"), sample_rate=22050, ) return save_id def ft2_ASRTTS_clean_hifi(audio_file, spk_name): spk = spks[spk_name] spembs = xvectors[spk] reg_text = transcriber(audio_file)["text"] speech, sr = torchaudio.load( audio_file, channels_first=True ) # Mono channel wav_tensor_spembs = ft2_text2speech_hifi( text=reg_text, speech=speech, spembs=spembs )["wav"] wav_numpy = wav_tensor_spembs.unsqueeze(1).to("cpu") sample_rate = 22050 save_id = ( "./wav/" + Path(audio_file).stem + "_fs2_hifi_" + spk_name + "_spkembs.wav" ) torchaudio.save( save_id, src=wav_tensor_spembs.unsqueeze(0).to("cpu"), sample_rate=22050, ) return save_id def transformer_ASRTTS_clean(audio_file, spk_name): spk = spks[spk_name] spembs = xvectors[spk] reg_text = transcriber(audio_file)["text"] speech, sr = torchaudio.load( audio_file, channels_first=True ) # Mono channel wav_tensor_spembs = transformer_text2speech( text=reg_text, speech=speech, spembs=spembs )["wav"] wav_numpy = wav_tensor_spembs.unsqueeze(1).to("cpu") sample_rate = 22050 save_id = ( "./wav/" + Path(audio_file).stem + "_transformer_" + spk_name + "_spkembs.wav" ) torchaudio.save( save_id, src=wav_tensor_spembs.unsqueeze(0).to("cpu"), sample_rate=22050, ) return save_id # def google_ASRTTS_clean(audio_file, spk_name): # spk = spks[spk_name] # spembs = xvectors[spk] # reg_text = transcriber(audio_file)["text"] # # pdb.set_trace() # synthesis_input = texttospeech.SynthesisInput(text=reg_text) # voice = texttospeech.VoiceSelectionParams( # language_code="en-US", ssml_gender=texttospeech.SsmlVoiceGender.NEUTRAL # ) # audio_config = texttospeech.AudioConfig( # audio_encoding=texttospeech.AudioEncoding.MP3 # ) # response = Google_TTS_client.synthesize_speech( # input=synthesis_input, voice=voice, audio_config=audio_config # ) # save_id = ( # "./wav/" + Path(audio_file).stem + "_google_" + spk_name + "_spkembs.wav" # ) # with open(save_id, "wb") as out_file: # out_file.write(response.audio_content) # return save_id reference_textbox = gr.Textbox( value="", placeholder="Input reference here", label="Reference", ) recognization_textbox = gr.Textbox( value="", placeholder="Output recognization here", label="recognization_textbox", ) speaker_option = gr.Radio(choices=spk_names, label="Speaker") input_audio = gr.Audio( source="upload", type="filepath", label="Audio_to_Evaluate" ) output_audio = gr.Audio( source="upload", file="filepath", label="Synthesized Audio" ) examples = [ ["./samples/001.wav", "M1", ""], ["./samples/002.wav", "M2", ""], ["./samples/003.wav", "F1", ""], ["./samples/004.wav", "F2", ""], ] def change_audiobox(choice): if choice == "upload": input_audio = gr.Audio(source="upload", visible=True) elif choice == "microphone": input_audio = gr.Audio(source="microphone", visible=True) else: input_audio = gr.Audio(visible=False) return input_audio def show_icon(choice): if choice == "Male1": spk_icon = gr.Image.update(value="speaker_icons/male1.png", visible=True) elif choice == "Male2": spk_icon = gr.Image.update(value="speaker_icons/male2.png", visible=True) elif choice == "Male3": spk_icon = gr.Image.update(value="speaker_icons/male3.png", visible=True) elif choice == "Female1": spk_icon = gr.Image.update(value="speaker_icons/female1.png", visible=True) elif choice == "Female2": spk_icon = gr.Image.update(value="speaker_icons/female2.png", visible=True) elif choice == "Female3": spk_icon = gr.Image.update(value="speaker_icons/female3.png", visible=True) return spk_icon def get_download_file(audio_file=None): if audio_file == None: output_audio_file = gr.File.update(visible=False) else: output_audio_file = gr.File.update(visible=True) return output_audio_file def download_file(audio_file): return gr.File(value=audio_file) # pdb.set_trace() with gr.Blocks( analytics_enabled=False, css=".gradio-container {background-color: #78BD91}", ) as demo: # Public Version with gr.Tab("Open Version"): with gr.Column(elem_id="Column"): input_format = gr.Radio( choices=["microphone", "upload"], label="Choose your input format", elem_id="input_format" ) input_audio = gr.Audio( source="microphone", type="filepath", label="Input Audio", interactive=True, visible=False, elem_id="input_audio" ) input_format.change( fn=change_audiobox, inputs=input_format, outputs=input_audio ) speaker_option = gr.Radio(choices=spk_names, value="Male1", label="Choose your voice profile") spk_icon = gr.Image(value="speaker_icons/male1.png", type="filepath", image_mode="RGB", source="upload", shape=[50, 50], interactive=True, visible=True) speaker_option.change( fn=show_icon, inputs=speaker_option, outputs=spk_icon ) b = gr.Button("Convert") output_audio = gr.Audio( source="upload", file="filepath", label="Converted Audio", interactive=False ) b.click( ASRTTS_clean, inputs=[input_audio, speaker_option], outputs=output_audio, api_name="convert" ) # # Tab selection: # with gr.Tab("Test Version: Multi TTS model"): # with gr.Column(elem_id="Column"): # input_format = gr.Radio( # choices=["microphone", "upload"], label="Choose your input format", elem_id="input_format" # ) # input_audio = gr.Audio( # source="microphone", # type="filepath", # label="Input Audio", # interactive=True, # visible=False, # elem_id="input_audio" # ) # input_format.change( # fn=change_audiobox, inputs=input_format, outputs=input_audio # ) # speaker_option = gr.Radio(choices=spk_names, value="Male1", label="Choose your voice profile") # spk_icon = gr.Image(value="speaker_icons/male1.png", # type="filepath", # image_mode="RGB", # source="upload", # shape=[50, 50], # interactive=True, # visible=True) # speaker_option.change( # fn=show_icon, inputs=speaker_option, outputs=spk_icon # ) # with gr.Column(): # with gr.Row(): # b2 = gr.Button("Convert") # output_audio = gr.Audio( # source="upload", file="filepath", label="Converted Audio", interactive=False # ) # b2.click( # ASRTTS_clean, # inputs=[input_audio, speaker_option], # outputs=output_audio, # api_name="convert_" # ) # with gr.Row(): # # Fastspeech2 + PWG [under construction] # b_ft2 = gr.Button("Convert_fastspeech2") # output_audio_ft2= gr.Audio( # source="upload", file="filepath", label="Converted Audio", interactive=False # ) # b_ft2.click( # ft2_ASRTTS_clean, # inputs=[input_audio, speaker_option], # outputs=output_audio_ft2, # api_name="convert_ft2" # ) # with gr.Row(): # # Fastspeech2 + hifigan [under construction] # b_ft2_hifi = gr.Button("Convert_fastspeech2+HifiGAN") # output_audio_ft2_hifi= gr.Audio( # source="upload", file="filepath", label="Converted Audio", interactive=False # ) # b_ft2_hifi.click( # ft2_ASRTTS_clean_hifi, # inputs=[input_audio, speaker_option], # outputs=output_audio_ft2_hifi, # api_name="convert_ft2_hifi" # ) # with gr.Row(): # # transformer [TODO] # b_transformer = gr.Button("Convert_transformer") # output_audio_transformer= gr.Audio( # source="upload", file="filepath", label="Converted Audio", interactive=False # ) # b_transformer.click( # transformer_ASRTTS_clean, # inputs=[input_audio, speaker_option], # outputs=output_audio_transformer, # api_name="convert_trans" # ) # google tts [TODO] # b_google = gr.Button("Convert_googleTTS") # output_audio_google= gr.Audio( # source="upload", file="filepath", label="Converted Audio", interactive=False # ) # b_google.click( # google_ASRTTS_clean, # inputs=[input_audio, speaker_option], # outputs=output_audio_google, # api_name="convert" # ) demo.launch(share=False)