training_config.json

{
    "project_name": "vits_kbd",
    "push_to_hub": true,
    "hub_model_id": "mms_finetune_kbd_murat",
    "overwrite_output_dir": false,
    "output_dir": "./tmp/vits_kbd_finetuned_che_model",
  
    "dataset_name": "anzorq/kbd_speech",
    "audio_column_name": "audio",
    "text_column_name": "transcription",
    "train_split_name": "train",
    "eval_split_name": "train",
    "speaker_id_column_name": "speaker_id",
    "override_speaker_embeddings": true,
    "filter_on_speaker_id": 9769125,

    "full_generation_sample_text": "укъэмыващэрэт тӏэ уэ щӏалэ цӏыкӏур! - жиӏащ лӏыжьым",
  
    "max_duration_in_seconds": 20,
    "min_duration_in_seconds": 1.0,
    "max_tokens_length": 500,

    "do_lower_case": true,
  
    "model_name_or_path": "anzorq/mms-tts-kbd-discriminator",
  
    "preprocessing_num_workers": 4,
  
    "do_train": true,
    "max_steps": 5100,
    "gradient_accumulation_steps": 1,
    "gradient_checkpointing": false,
    "per_device_train_batch_size": 16,
    "learning_rate": 1e-4,
    "adam_beta1": 0.8,
    "adam_beta2": 0.99,
    "warmup_ratio": 0.01,
    "group_by_length": false,
  
    "do_eval": true,
    "eval_steps": 100,
    "per_device_eval_batch_size": 16,
    "max_eval_samples": 100,
    "do_step_schedule_per_epoch": true,
  
    "weight_disc": 3,
    "weight_fmaps": 1,
    "weight_gen": 1,
    "weight_kl": 1.5,
    "weight_duration": 1,
    "weight_mel": 35,
  
    "fp16": true,
    "seed": 456
  }