Update VCTK recipes

TTS/tts/datasets/formatters.py

@@ -289,7 +289,7 @@ def brspeech(root_path, meta_file, ignored_speakers=None):
     return items
 
 
-def vctk(root_path, meta_files=None, wavs_path="wav22", mic="mic2", ignored_speakers=None):
+def vctk(root_path, meta_files=None, wavs_path="wav48_silence_trimmed", mic="mic2", ignored_speakers=None):
     """https://datashare.ed.ac.uk/bitstream/handle/10283/3443/VCTK-Corpus-0.92.zip"""
     file_ext = 'flac'
     test_speakers = meta_files

recipes/ljspeech/speedy_speech/train_speedy_speech.py

@@ -68,12 +68,6 @@ tokenizer, config = TTSTokenizer.init_from_config(config)
 # Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
-# init audio processor
-ap = AudioProcessor(**config.audio.to_dict())
-
-# load training samples
-train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
-
 # init model
 model = ForwardTTS(config, ap, tokenizer)
 

recipes/vctk/fast_pitch/train_fast_pitch.py

@@ -6,6 +6,7 @@ from TTS.tts.configs.fast_pitch_config import FastPitchConfig
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.forward_tts import ForwardTTS
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -32,6 +33,7 @@ config = FastPitchConfig(
     num_loader_workers=8,
     num_eval_loader_workers=4,
     compute_input_seq_cache=True,
+    precompute_num_workers=4,
     compute_f0=True,
     f0_cache_path=os.path.join(output_path, "f0_cache"),
     run_eval=True,
@@ -39,23 +41,35 @@ config = FastPitchConfig(
     epochs=1000,
     text_cleaner="english_cleaners",
     use_phonemes=True,
-    use_espeak_phonemes=False,
     phoneme_language="en-us",
     phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
     print_step=50,
     print_eval=False,
     mixed_precision=False,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    max_seq_len=500000,
+    max_text_len=500,
+    min_audio_len=0,
+    max_audio_len=500000,
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,
 )
 
-# init audio processor
+# INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio)
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -65,16 +79,15 @@ speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 config.model_args.num_speakers = speaker_manager.num_speakers
 
 # init model
-model = ForwardTTS(config, speaker_manager)
+model = ForwardTTS(config, ap, tokenizer, speaker_manager=speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()
+

recipes/vctk/fast_speech/train_fast_speech.py

@@ -6,6 +6,7 @@ from TTS.tts.configs.fast_speech_config import FastSpeechConfig
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.forward_tts import ForwardTTS
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -25,37 +26,48 @@ audio_config = BaseAudioConfig(
 )
 
 config = FastSpeechConfig(
-    run_name="fast_pitch_ljspeech",
+    run_name="fast_speech_vctk",
     audio=audio_config,
     batch_size=32,
     eval_batch_size=16,
     num_loader_workers=8,
     num_eval_loader_workers=4,
     compute_input_seq_cache=True,
-    compute_f0=True,
+    precompute_num_workers=4,
-    f0_cache_path=os.path.join(output_path, "f0_cache"),
     run_eval=True,
     test_delay_epochs=-1,
     epochs=1000,
     text_cleaner="english_cleaners",
     use_phonemes=True,
-    use_espeak_phonemes=False,
     phoneme_language="en-us",
     phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
     print_step=50,
     print_eval=False,
     mixed_precision=False,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    max_seq_len=500000,
+    max_text_len=500,
+    min_audio_len=0,
+    max_audio_len=500000,
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,
 )
 
-# init audio processor
+## INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio)
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -65,16 +77,14 @@ speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 config.model_args.num_speakers = speaker_manager.num_speakers
 
 # init model
-model = ForwardTTS(config, speaker_manager)
+model = ForwardTTS(config, ap, tokenizer, speaker_manager=speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()

recipes/vctk/glow_tts/train_glow_tts.py

@@ -7,6 +7,7 @@ from TTS.tts.configs.shared_configs import BaseDatasetConfig
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.glow_tts import GlowTTS
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 # set experiment paths
@@ -32,6 +33,7 @@ config = GlowTTSConfig(
     eval_batch_size=16,
     num_loader_workers=4,
     num_eval_loader_workers=4,
+    precompute_num_workers=4,
     run_eval=True,
     test_delay_epochs=-1,
     epochs=1000,
@@ -45,12 +47,27 @@ config = GlowTTSConfig(
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,
+    min_text_len=0,
+    max_text_len=500,
+    min_audio_len=0,
+    max_audio_len=500000,
 )
 
-# init audio processor
+# INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio.to_dict())
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -60,16 +77,14 @@ speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 config.num_speakers = speaker_manager.num_speakers
 
 # init model
-model = GlowTTS(config, speaker_manager)
+model = GlowTTS(config, ap, tokenizer, speaker_manager=speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()

recipes/vctk/speedy_speech/train_speedy_speech.py

@@ -6,6 +6,7 @@ from TTS.tts.configs.speedy_speech_config import SpeedySpeechConfig
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.forward_tts import ForwardTTS
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -32,30 +33,41 @@ config = SpeedySpeechConfig(
     num_loader_workers=8,
     num_eval_loader_workers=4,
     compute_input_seq_cache=True,
-    compute_f0=True,
+    precompute_num_workers=4,
-    f0_cache_path=os.path.join(output_path, "f0_cache"),
     run_eval=True,
     test_delay_epochs=-1,
     epochs=1000,
     text_cleaner="english_cleaners",
     use_phonemes=True,
-    use_espeak_phonemes=False,
     phoneme_language="en-us",
     phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
     print_step=50,
     print_eval=False,
     mixed_precision=False,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    max_seq_len=500000,
+    max_text_len=500,
+    min_audio_len=0,
+    max_audio_len=500000,
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,
 )
 
-# init audio processor
+# INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio)
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -65,16 +77,14 @@ speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 config.model_args.num_speakers = speaker_manager.num_speakers
 
 # init model
-model = ForwardTTS(config, speaker_manager)
+model = ForwardTTS(config, ap, tokenizer, speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()

recipes/vctk/tacotron-DDC/train_tacotron-DDC.py

@@ -7,6 +7,7 @@ from TTS.tts.configs.tacotron_config import TacotronConfig
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.tacotron import Tacotron
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -32,6 +33,7 @@ config = TacotronConfig(  # This is the config that is saved for the future use
     eval_batch_size=16,
     num_loader_workers=4,
     num_eval_loader_workers=4,
+    precompute_num_workers=4,
     run_eval=True,
     test_delay_epochs=-1,
     r=6,
@@ -45,18 +47,30 @@ config = TacotronConfig(  # This is the config that is saved for the future use
     print_step=25,
     print_eval=False,
     mixed_precision=True,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    min_seq_len=0,
+    max_text_len=500,
-    max_seq_len=44000 * 10,  # 44k is the original sampling rate before resampling, corresponds to 10 seconds of audio
+    min_audio_len=0,
+    max_audio_len=44000 * 10,  # 44k is the original sampling rate before resampling, corresponds to 10 seconds of audio
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,  # set this to enable multi-sepeaker training
 )
 
-# init audio processor
+## INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio.to_dict())
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -65,16 +79,14 @@ speaker_manager = SpeakerManager()
 speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 
 # init model
-model = Tacotron(config, speaker_manager)
+model = Tacotron(config, ap, tokenizer, speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()

recipes/vctk/tacotron2-DDC/train_tacotron2-ddc.py

@@ -7,6 +7,7 @@ from TTS.tts.configs.tacotron2_config import Tacotron2Config
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.tacotron2 import Tacotron2
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -44,9 +45,10 @@ config = Tacotron2Config(  # This is the config that is saved for the future use
     print_step=150,
     print_eval=False,
     mixed_precision=True,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    min_seq_len=14800,
+    max_text_len=500,
-    max_seq_len=22050 * 10,  # 44k is the original sampling rate before resampling, corresponds to 10 seconds of audio
+    min_audio_len=0,
+    max_audio_len=44000 * 10,
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,  # set this to enable multi-sepeaker training
@@ -60,10 +62,21 @@ config = Tacotron2Config(  # This is the config that is saved for the future use
     lr=3e-5,
 )
 
-# init audio processor
+# INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio.to_dict())
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -72,16 +85,14 @@ speaker_manager = SpeakerManager()
 speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 
 # init model
-model = Tacotron2(config, speaker_manager)
+model = Tacotron2(config, ap, tokenizer, speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()

recipes/vctk/tacotron2/train_tacotron2.py

@@ -7,6 +7,7 @@ from TTS.tts.configs.tacotron2_config import Tacotron2Config
 from TTS.tts.datasets import load_tts_samples
 from TTS.tts.models.tacotron2 import Tacotron2
 from TTS.tts.utils.speakers import SpeakerManager
+from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.utils.audio import AudioProcessor
 
 output_path = os.path.dirname(os.path.abspath(__file__))
@@ -44,9 +45,10 @@ config = Tacotron2Config(  # This is the config that is saved for the future use
     print_step=150,
     print_eval=False,
     mixed_precision=True,
-    sort_by_audio_len=True,
+    min_text_len=0,
-    min_seq_len=14800,
+    max_text_len=500,
-    max_seq_len=22050 * 10,  # 44k is the original sampling rate before resampling, corresponds to 10 seconds of audio
+    min_audio_len=0,
+    max_audio_len=44000 * 10,
     output_path=output_path,
     datasets=[dataset_config],
     use_speaker_embedding=True,  # set this to enable multi-sepeaker training
@@ -60,10 +62,21 @@ config = Tacotron2Config(  # This is the config that is saved for the future use
     lr=3e-5,
 )
 
-# init audio processor
+## INITIALIZE THE AUDIO PROCESSOR
-ap = AudioProcessor(**config.audio.to_dict())
+# Audio processor is used for feature extraction and audio I/O.
+# It mainly serves to the dataloader and the training loggers.
+ap = AudioProcessor.init_from_config(config)
 
-# load training samples
+# INITIALIZE THE TOKENIZER
+# Tokenizer is used to convert text to sequences of token IDs.
+# If characters are not defined in the config, default characters are passed to the config
+tokenizer, config = TTSTokenizer.init_from_config(config)
+
+# LOAD DATA SAMPLES
+# Each sample is a list of ```[text, audio_file_path, speaker_name]```
+# You can define your custom sample loader returning the list of samples.
+# Or define your custom formatter and pass it to the `load_tts_samples`.
+# Check `TTS.tts.datasets.load_tts_samples` for more details.
 train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)
 
 # init speaker manager for multi-speaker training
@@ -72,16 +85,14 @@ speaker_manager = SpeakerManager()
 speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)
 
 # init model
-model = Tacotron2(config, speaker_manager)
+model = Tacotron2(config, ap, tokenizer, speaker_manager)
 
-# init the trainer and 🚀
+# INITIALIZE THE TRAINER
+# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
+# distributed training, etc.
 trainer = Trainer(
-    TrainingArgs(),
+    TrainingArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
-    config,
-    output_path,
-    model=model,
-    train_samples=train_samples,
-    eval_samples=eval_samples,
-    training_assets={"audio_processor": ap},
 )
+
+# AND... 3,2,1... 🚀
 trainer.fit()