config/base.json

{
  "supported_model_type": [
    "GANVocoder",
    "Fastspeech2",
    "DiffSVC",
    "Transformer",
    "EDM",
    "CD"
  ],
  "task_type": "",
  "dataset": [],
  "use_custom_dataset": [],
  "preprocess": {
    "phone_extractor": "espeak", // "espeak, pypinyin, pypinyin_initials_finals, lexicon"
    // trim audio silence
    "data_augment": false,
    "trim_silence": false,
    "num_silent_frames": 8,
    "trim_fft_size": 512, // fft size used in trimming
    "trim_hop_size": 128, // hop size used in trimming
    "trim_top_db": 30, // top db used in trimming sensitive to each dataset
    // acoustic features
    "extract_mel": false,
    "mel_extract_mode": "",
    "extract_linear_spec": false,
    "extract_mcep": false,
    "extract_pitch": false,
    "extract_acoustic_token": false,
    "pitch_remove_outlier": false,
    "extract_uv": false,
    "pitch_norm": false,
    "extract_audio": false,
    "extract_label": false,
    "pitch_extractor": "parselmouth", // pyin, dio, pyworld, pyreaper, parselmouth, CWT (Continuous Wavelet Transform)
    "extract_energy": false,
    "energy_remove_outlier": false,
    "energy_norm": false,
    "energy_extract_mode": "from_mel",
    "extract_duration": false,
    "extract_amplitude_phase": false,
    "mel_min_max_norm": false,
    // lingusitic features
    "extract_phone": false,
    "lexicon_path": "./text/lexicon/librispeech-lexicon.txt",
    // content features
    "extract_whisper_feature": false,
    "extract_contentvec_feature": false,
    "extract_mert_feature": false,
    "extract_wenet_feature": false,
    // Settings for data preprocessing
    "n_mel": 80,
    "win_size": 480,
    "hop_size": 120,
    "sample_rate": 24000,
    "n_fft": 1024,
    "fmin": 0,
    "fmax": 12000,
    "min_level_db": -115,
    "ref_level_db": 20,
    "bits": 8,
    // Directory names of processed data or extracted features
    "processed_dir": "processed_data",
    "trimmed_wav_dir": "trimmed_wavs", // directory name of silence trimed wav
    "raw_data": "raw_data",
    "phone_dir": "phones",
    "wav_dir": "wavs", // directory name of processed wav (such as downsampled waveform)
    "audio_dir": "audios",
    "log_amplitude_dir": "log_amplitudes",
    "phase_dir": "phases",
    "real_dir": "reals",
    "imaginary_dir": "imaginarys",
    "label_dir": "labels",
    "linear_dir": "linears",
    "mel_dir": "mels", // directory name of extraced mel features
    "mcep_dir": "mcep", // directory name of extraced mcep features
    "dur_dir": "durs",
    "symbols_dict": "symbols.dict",
    "lab_dir": "labs", // directory name of extraced label features
    "wenet_dir": "wenet", // directory name of extraced wenet features
    "contentvec_dir": "contentvec", // directory name of extraced wenet features
    "pitch_dir": "pitches", // directory name of extraced pitch features
    "energy_dir": "energys", // directory name of extracted energy features
    "phone_pitch_dir": "phone_pitches", // directory name of extraced pitch features
    "phone_energy_dir": "phone_energys", // directory name of extracted energy features
    "uv_dir": "uvs", // directory name of extracted unvoiced features
    "duration_dir": "duration", // ground-truth duration file
    "phone_seq_file": "phone_seq_file", // phoneme sequence file
    "file_lst": "file.lst",
    "train_file": "train.json", // training set, the json file contains detailed information about the dataset, including dataset name, utterance id, duration of the utterance
    "valid_file": "valid.json", // validattion set
    "spk2id": "spk2id.json", // used for multi-speaker dataset
    "utt2spk": "utt2spk", // used for multi-speaker dataset
    "emo2id": "emo2id.json", // used for multi-emotion dataset
    "utt2emo": "utt2emo", // used for multi-emotion dataset
    // Features used for model training
    "use_text": false,
    "use_phone": false,
    "use_phn_seq": false,
    "use_lab": false,
    "use_linear": false,
    "use_mel": false,
    "use_min_max_norm_mel": false,
    "use_wav": false,
    "use_phone_pitch": false,
    "use_log_scale_pitch": false,
    "use_phone_energy": false,
    "use_phone_duration": false,
    "use_log_scale_energy": false,
    "use_wenet": false,
    "use_dur": false,
    "use_spkid": false, // True: use speaker id for multi-speaker dataset
    "use_emoid": false, // True: use emotion id for multi-emotion dataset
    "use_frame_pitch": false,
    "use_uv": false,
    "use_frame_energy": false,
    "use_frame_duration": false,
    "use_audio": false,
    "use_label": false,
    "use_one_hot": false,
    "use_amplitude_phase": false,
    "align_mel_duration": false
  },
  "train": {
    "ddp": true,
    "batch_size": 16,
    "max_steps": 1000000,
    // Trackers
    "tracker": [
      "tensorboard"
      // "wandb",
      // "cometml",
      // "mlflow",
    ],
    "max_epoch": -1,
    // -1 means no limit
    "save_checkpoint_stride": [
      5,
      20
    ],
    // unit is epoch
    "keep_last": [
      3,
      -1
    ],
    // -1 means infinite, if one number will broadcast
    "run_eval": [
      false,
      true
    ],
    // if one number will broadcast
    // Fix the random seed
    "random_seed": 10086,
    // Optimizer
    "optimizer": "AdamW",
    "adamw": {
      "lr": 4.0e-4
      // nn model lr
    },
    // LR Scheduler
    "scheduler": "ReduceLROnPlateau",
    "reducelronplateau": {
      "factor": 0.8,
      "patience": 10,
      // unit is epoch
      "min_lr": 1.0e-4
    },
    // Batchsampler
    "sampler": {
      "holistic_shuffle": true,
      "drop_last": true
    },
    // Dataloader
    "dataloader": {
      "num_worker": 32,
      "pin_memory": true
    },
    "gradient_accumulation_step": 1,
    "total_training_steps": 50000,
    "save_summary_steps": 500,
    "save_checkpoints_steps": 10000,
    "valid_interval": 10000,
    "keep_checkpoint_max": 5,
    "multi_speaker_training": false // True: train multi-speaker model; False: training single-speaker model;
  }
}