I have a few questions regarding the implementation of Bark scale filter banks for the calculation of bark frequency cepstral coefficients.
1) Which filter shape is most commonly used and why?
- Triangular band pass similar to the mfcc's as seen here:- A Comparative Study of Filter Bank Spacing for Speech Recognition
- Trapazoidal type shape as seen here:- Performance Evaluation of Front-end Processing for Speech Recognition Systems and here:- Auditory Models for Speech Analysis
- 4th order elliptical bandpass filters used by the Mir toolbox
2) How should the filters be laid out?
- No overlap with centre frequencies 1 bark apart?
- Half a bark overlap and a length of 1 bark?
- Linearly divided on the bark scale between a predefined frequency range and number of filters? If so how is a suitable number of filters chosen e.g. 40Hz to fs/2
Thanks for any help provided, I've become pretty confused from reading the literature