sp500_log_mel_params¶
Front-end parameters.
Fill this with sp500_log_mel_params_init(); the defaults are the Whisper 80-mel front end. Qwen3-ASR and Whisper large-v3 use n_mels = 128. Only n_samples has no useful default and must be set.
Attributes¶
| Member | Description |
|---|---|
n_samples |
|
sample_rate |
|
n_fft |
|
hop_length |
|
n_mels |
|
min_frequency |
|
max_frequency |
|
normalize |
|
reserved0 |
|
mel_filters |
|
mel_filters_count |
|
reserved |
n_samples¶
number of input samples (required)
sample_rate¶
sampling rate in Hz; default 16000
n_fft¶
FFT size; default 400. Must be 16..2048 and factor into 2, 3 and 5 only
hop_length¶
hop in samples; default 160. Must be 1..65535
n_mels¶
mel bins; default 80. Must be 1..256
min_frequency¶
lower edge of the mel bank in Hz; default 0
max_frequency¶
upper edge of the mel bank in Hz; default 8000
normalize¶
1: apply the Whisper clamp and rescale. Default 1
reserved0¶
must be zero
mel_filters¶
Optional dense mel bank, [n_fft/2 + 1][n_mels] row-major - the same layout as Hugging Face mel_filters. NULL generates the slaney bank from the parameters above. Read only during the call; it does not need to outlive it.
mel_filters_count¶
Number of floats in mel_filters. Must be sp500_log_mel_num_freq(n_fft) * n_mels when mel_filters is set, and 0 when it is NULL.
reserved¶
must be zero