| { | |
| "feature_type": "log_mel_spectrogram", | |
| "sampling_rate": 16000, | |
| "mono": true, | |
| "n_mels": 80, | |
| "n_fft": 400, | |
| "win_length": 400, | |
| "hop_length": 160, | |
| "center": false, | |
| "f_min": 20.0, | |
| "f_max": 7600.0, | |
| "normalization": { | |
| "type": "per_utterance_per_frequency", | |
| "epsilon": 1e-05, | |
| "formula": "(mel - mean) / (std + 1e-5)", | |
| "std_correction": 1 | |
| }, | |
| "important_note": "For chunked inference, compute the complete utterance log-Mel spectrogram and normalize it before slicing it into inference chunks.", | |
| "power": 2.0, | |
| "window_fn": "hann", | |
| "normalized_stft": false, | |
| "mel_filterbank": { | |
| "norm": null, | |
| "mel_scale": "htk" | |
| }, | |
| "log_transform": { | |
| "type": "natural_log", | |
| "clamp_min": 1e-05 | |
| } | |
| } |