training_config.json

{
  "tokenizer_id": "zuhri025/TTS-Tokenizer-T",
  "model_config_id": "ekwek/Soprano-1.1-80M",
  "save_dir": "tts_weights_T2",
  "repo_name": "tts_weights_T2",
  "active_datasets": [
    "globe",
    "long_audio",
    "emilia",
    "gemini",
    "soda"
  ],
  "dataset_limits": {
    "gemini": 10000,
    "globe": 250000,
    "soda": 250000,
    "emilia": 80000,
    "long_audio": 70000
  },
  "max_samples": 100000,
  "device": "cuda",
  "num_workers": 2,
  "pin_memory": true,
  "compile_model": true,
  "use_mixed_precision": true,
  "seed": 1337,
  "max_steps": 20000,
  "batch_size": 2,
  "grad_accum_steps": 4,
  "seq_len": 1024,
  "max_lr": 0.001,
  "min_lr": 5e-05,
  "warmup_ratio": 0.1,
  "cooldown_ratio": 0.1,
  "betas": [
    0.9,
    0.95
  ],
  "weight_decay": 0.1,
  "grad_clip_norm": 1.0,
  "text_loss_weight": 0.5,
  "text_mode_ratio": 1.0,
  "phoneme_mode_ratio": 0.0,
  "mix_mode_ratio": 0.0,
  "unconditional_prob": 0.05,
  "voice_dropout_prob": 0.1,
  "word_mix_ratio": 0.5,
  "val_freq": 1500,
  "save_freq": 1500,
  "log_freq": 10,
  "val_steps": 20,
  "test_split_size": 0.01,
  "upload_tensorboard": true,
  "tensorboard_dir": "tensorboard_logs"
}