A Comparative Study of Filter Bank Spacing for Speech Recognition
Ben James Shannon, Kuldip K. Paliwal · 2003
Although Mel scale filter bank spacing is used extensively in Automatic Speech Recognition (ASR), it will be shown in this paper that it provides little benefit over other perceptually motivated frequency warping scales. An MFCC like feature based on the Bark scale is shown to yield similar performance in speech recognition experiments as MFCC. The performance of MFCC and BFCC features are also compared to Uniform Frequency Cepstral Coefficients (UFCC) where it is shown that neither the Mel or Bark scale provide significant advantage over a Uniform scale if training and test conditions are matched.