Comparing various MFCC features

In Figure 11.2, we show the acid test results for three MFCC chains: software/module_mel_ml.m, software/module_mel_bank.m with method='spa', and software/module_mel_bank.m with method='clt'. The best was ML, followed by SPA, then CLT. This means that for estimating the PDF of data that has a spectral structure consistent with MFCC, that software/module_mel_ml.m is superior. This does not necessarily mean it is better in, say, human speech applications.

Figure 11.2: Acid test performance of three types of MFCC feature chains for data generated according to the circular MFCC model.