training_config.json

{
  "audio_dir": "audio_data",
  "transcript_file": "audio_data/transcripts.txt",
  "metadata_file": "audio_data/metadata.json",
  "output_dir": "asr_model_output",
  "model_save_path": "asr_model_final",
  "base_model": "openai/whisper-tiny",
  "language": "tagalog",
  "task": "transcribe",
  "test_split_ratio": 0.15,
  "max_input_length_s": 30.0,
  "batch_size": 8,
  "grad_accum_steps": 2,
  "learning_rate": 1e-05,
  "warmup_steps": 100,
  "max_steps": 2000,
  "eval_steps": 200,
  "save_steps": 200,
  "fp16": true,
  "generation_max_length": 225,
  "num_beams": 5
}