Comparison between wavelet packet transform, Bark Wavelet & MFCC for robust speech recognition tasks

Although Wavelet Transformation has multi resolution properties, it is not optimized for speech recognition tasks. There are two major perspectives, the first approach is based on selection of similar frequencies in perceptual auditory scale using wavelet packet and the second one involves frequency aspects of continuous wavelet leading to Bark Wavelet. This paper shows that because ordinary wavelet packet transform is time variant, it´s filter bank has high overlapping and exact auditory band width in Bark scale cannot be achieved, this transform works weaker than perceptual representations for speech recognition, specially under noisy conditions. Mel Frequency Cepstral Coefficient is one of the famous methods using for speech recognition and is optimized for speech recognition. This paper shows that because time-frequency localization capability of bark wavelet transform together with its multi-resolution property makes it more suitable than Discrete Cosine transform, Bark wavelet works better than MFCC and wavelet packet in noisy conditions.