Motivation
Keyword-spotting and spectrogram models on MCUs (DS-CNN, the "Hello Edge" class of models) need a log-mel or MFCC front end. The analysis module already has the parts: real FFT, DCT and windowing. It has no mel filterbank or MFCC stage. This is neural-network-toobox-cpp roadmap item N15, which feeds N18/N19 (2D and depthwise-separable convolution).
Proposed scope
- Windowed real FFT → power spectrum → triangular mel filterbank (HTK or Slaney scale, with
NumMelBands, fMin and fMax as parameters) → log (with ε floor) → DCT-II (first NumCoefficients).
- Compile-time sizes, float-only, no heap, filterbank weights precomputable (constexpr where feasible).
- Tests: the mel/Hz conversion round-trip, filterbank partition of unity (Slaney normalisation), and one frame checked against a librosa/python_speech_features reference with the same parameters.
References
S. Davis, P. Mermelstein, IEEE Trans. ASSP 28(4), 1980.
Motivation
Keyword-spotting and spectrogram models on MCUs (DS-CNN, the "Hello Edge" class of models) need a log-mel or MFCC front end. The
analysismodule already has the parts: real FFT, DCT and windowing. It has no mel filterbank or MFCC stage. This is neural-network-toobox-cpp roadmap item N15, which feeds N18/N19 (2D and depthwise-separable convolution).Proposed scope
NumMelBands,fMinandfMaxas parameters) → log (with ε floor) → DCT-II (firstNumCoefficients).References
S. Davis, P. Mermelstein, IEEE Trans. ASSP 28(4), 1980.