Module features 

Kaldi-style FBANK (log-mel filterbank) feature extraction.

Produces the 80-dim features that the wespeaker speaker-embedding ONNX model in crate::voice::speaker consumes. Parameters match sherpa-onnx’s kaldi-native-fbank defaults for that model:

• 16 kHz mono input
• 25 ms window (400 samples) / 10 ms hop (160 samples)
• Pre-emphasis 0.97
• Hamming window
• 80 mel bins, low-freq 20 Hz, high-freq 8000 Hz
• Slaney/Kaldi mel scale: 1127 · ln(1 + f/700)
• Log applied to mel energies (ln(power + 1e-10))
• Cepstral mean normalisation across all frames of the supplied window

The implementation is pure-Rust DSP; the only non-std dep is rustfft for the 512-point real FFT.

Constants

FFT_SIZE

FFT size: smallest power of two ≥ frame length (400 → 512).

FRAME_LENGTH_MS

Frame length in milliseconds.

FRAME_SHIFT_MS

Frame shift (hop) in milliseconds.

HIGH_FREQ_HZ

Mel-filter high-frequency cutoff (Hz). Half the sample rate.

LOW_FREQ_HZ

Mel-filter low-frequency cutoff (Hz). Kaldi default.

NUM_MEL_BINS

Number of mel bins (also the feature dimension per frame).

PREEMPHASIS

Pre-emphasis coefficient — boost high frequencies before windowing.

SAMPLE_RATE

Target sample rate: wespeaker is trained on 16 kHz audio.

Functions

build_mel_filterbank

Builds the triangular mel filterbank as a dense [num_bins][fft_size/2 + 1] matrix indexed [mel_bin][fft_bin].

compute_fbank

Computes 80-dim Kaldi-style FBANK features for the supplied 16 kHz mono floating-point PCM window. Returns one feature row per frame, already cepstral-mean-normalised across the window.