| 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 | |