Delta MFCC and PCEN: Feature Extraction for Enhanced Speech Recognition

In our system, we use a combination of delta MFCC and PCEN as features. Delta MFCC involves performing cepstrum analysis on the Mel-spectrogram to derive the Mel-scale Frequency Cepstral Coefficients, which are then mixed with the original MFCC to obtain delta MFCC. On the other hand, PCEN uses FFT or Fbank features to normalize each channel and minimize the impact of input signal amplitude changes on recognition results.