finetune_mms_kor.json

{
    "project_name": "mms_korean-finetuning",
    "push_to_hub": true,
    "hub_model_id": "ylacombe/mms-tts-kor-finetuned",
    "report_to": ["wandb"],
    "overwrite_output_dir": true,
    "output_dir": "./tmp/vits_finetuned_kor",

    "dataset_name": "PolyAI/minds14",    
    "dataset_config_name": "ko-KR", 
    "audio_column_name": "audio", 
    "text_column_name":"transcription",
    "train_split_name": "train",
    "eval_split_name": "train",

    "full_generation_sample_text": "그는 괜찮은 척하려고 애쓰는 것 같았다.",
    
    "max_duration_in_seconds": 20,
    "min_duration_in_seconds": 1.0,
    "max_tokens_length": 500,

    "model_name_or_path": "ylacombe/mms-tts-kor-train",


    "preprocessing_num_workers": 4,

    "do_train": true,
    "num_train_epochs": 200,
    "max_train_samples": 150,
    "gradient_accumulation_steps": 1,
    "gradient_checkpointing": false,
    "per_device_train_batch_size": 16,
    "learning_rate": 2e-5,
    "adam_beta1": 0.8,
    "adam_beta2": 0.99,
    "warmup_ratio": 0.01,
    "group_by_length": false,

    "do_eval": true, 
    "eval_steps": 50,
    "per_device_eval_batch_size": 16,
    "max_eval_samples": 25, 
    "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,

    "uroman_path": "/home/yoach/uroman"
}