| { | |
| "name": "default_config", | |
| "n_gpu": 1, | |
| "text_encoder": { | |
| "type": "CTCCharTextEncoder", | |
| "args": { | |
| "kenlm_model_path": "hw_asr/text_encoder/3-gram.arpa", | |
| "unigrams_path": "hw_asr/text_encoder/librispeech-vocab.txt" | |
| } | |
| }, | |
| "preprocessing": { | |
| "sr": 16000, | |
| "spectrogram": { | |
| "type": "MelSpectrogram", | |
| "args": { | |
| "n_mels": 256 | |
| } | |
| }, | |
| "log_spec": true | |
| }, | |
| "augmentations": { | |
| "random_apply_p": 0.6, | |
| "wave": [ | |
| {"type": "AddColoredNoise", "args": {"p": 1, "sample_rate": 16000}}, | |
| {"type": "Gain", "args": {"p": 0.8, "sample_rate": 16000}}, | |
| {"type": "HighPassFilter", "args": {"p": 0, "sample_rate": 16000}}, | |
| {"type": "LowPassFilter", "args": {"p": 0, "sample_rate": 16000}}, | |
| {"type": "PitchShift", "args": {"p": 0.8, "min_transpose_semitones": -2, "max_transpose_semitones": 2, "sample_rate": 16000}}, | |
| {"type": "PolarityInversion", "args": {"p": 0.8, "sample_rate": 16000}}, | |
| {"type": "Shift", "args": {"p": 0.8, "sample_rate": 16000}} | |
| ], | |
| "spectrogram": [] | |
| }, | |
| "arch": { | |
| "type": "DeepSpeech2Model", | |
| "args": { | |
| "n_feats": 256, | |
| "n_rnn_layers": 5, | |
| "rnn_hidden_size": 512, | |
| "rnn_dropout": 0.2 | |
| } | |
| }, | |
| "data": { | |
| "train": { | |
| "batch_size": 128, | |
| "num_workers": 4, | |
| "datasets": [ | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "train-clean-100", | |
| "max_audio_length": 40.0, | |
| "max_text_length": 400 | |
| } | |
| }, | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "train-clean-360", | |
| "max_audio_length": 40.0, | |
| "max_text_length": 400 | |
| } | |
| }, | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "train-other-500", | |
| "max_audio_length": 40.0, | |
| "max_text_length": 400 | |
| } | |
| } | |
| ] | |
| }, | |
| "val": { | |
| "batch_size": 64, | |
| "num_workers": 4, | |
| "datasets": [ | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "dev-clean" | |
| } | |
| } | |
| ] | |
| }, | |
| "test-other": { | |
| "batch_size": 64, | |
| "num_workers": 4, | |
| "datasets": [ | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "test-other" | |
| } | |
| } | |
| ] | |
| }, | |
| "test-clean": { | |
| "batch_size": 64, | |
| "num_workers": 4, | |
| "datasets": [ | |
| { | |
| "type": "LibrispeechDataset", | |
| "args": { | |
| "part": "test-clean" | |
| } | |
| } | |
| ] | |
| } | |
| }, | |
| "optimizer": { | |
| "type": "AdamW", | |
| "args": { | |
| "lr": 5e-4, | |
| "weight_decay": 1e-3 | |
| } | |
| }, | |
| "loss": { | |
| "type": "CTCLoss", | |
| "args": {} | |
| }, | |
| "metrics": [ | |
| { | |
| "type": "ArgmaxWERMetric", | |
| "args": { | |
| "name": "WER (argmax)" | |
| } | |
| }, | |
| { | |
| "type": "ArgmaxCERMetric", | |
| "args": { | |
| "name": "CER (argmax)" | |
| } | |
| }, | |
| { | |
| "type": "BeamSearchWERMetric", | |
| "args": { | |
| "beam_size": 4, | |
| "name": "WER (beam search)" | |
| } | |
| }, | |
| { | |
| "type": "BeamSearchCERMetric", | |
| "args": { | |
| "beam_size": 4, | |
| "name": "CER (beam search)" | |
| } | |
| }, | |
| { | |
| "type": "LanguageModelWERMetric", | |
| "args": { | |
| "name": "WER (LM)" | |
| } | |
| }, | |
| { | |
| "type": "LanguageModelCERMetric", | |
| "args": { | |
| "name": "CER (LM)" | |
| } | |
| } | |
| ], | |
| "lr_scheduler": { | |
| "type": "OneCycleLR", | |
| "args": { | |
| "steps_per_epoch": 1000, | |
| "epochs": 50, | |
| "anneal_strategy": "cos", | |
| "max_lr": 5e-4, | |
| "pct_start": 0.1 | |
| } | |
| }, | |
| "trainer": { | |
| "epochs": 50, | |
| "save_dir": "saved/", | |
| "save_period": 5, | |
| "verbosity": 2, | |
| "monitor": "min val_loss", | |
| "early_stop": 100, | |
| "visualize": "wandb", | |
| "wandb_project": "asr_project", | |
| "len_epoch": 1000, | |
| "grad_norm_clip": 10 | |
| } | |
| } | |