input:
  shape: [1, 80, 3000]
  dtype: float32
  description: "Log-Mel spectrogram (batch=1, mel_bins=80, time_frames=3000 = 30s at 16kHz)"
  preprocessing:
    - load_audio_as_waveform_16khz
    - apply_log_mel_spectrogram:
        n_mels: 80
        hop_length: 160
        n_fft: 400
        sample_rate: 16000
        duration_s: 30
    - pad_or_trim_to_3000_frames

output:
  format: "Token ID sequence (autoregressive generation)"
  postprocessing:
    - greedy_decode_with_whisper_tokenizer
    - skip_special_tokens: true