machinehearing

Preprocessing

• A-weighting
• Log transform
• Harmonic-percussive-residual source separation. Especially for music.

• Per-channel energy normalization (PCEN). Static version exists as librosa.pen. Can also be learned as a neural network layer, see arXiv:1607.05666v1 Per-Channel Energy Normalization: Why and How.

• Whitening. Eg PCA. Removes redundancies in spectrogram. For each frame in spectogram

Normalization

• Cepstral Mean Normalisation (CMN): subtract the average feature value from each feature, so each feature has a mean value of 0. makes features robust to some linear filtering of the signal (channel variation).
• Cepstral Variance Normalisation (CVN): Divide feature vector by standard deviation of feature vectors, so each feature vector element has a variance of 1.
• For real-time, need to compute a moving average.
• RMS normalization
• Gaussianization, mapping to Gaussian distribution