Best wavelet decomposition channel determination for speech processing application using two-way ANOVA

This study aims to determine the best channel of wavelet decomposition’s result as features on speaker recognition application or automatic speech recognition. We process fundamental (F0) and formant (F1-F4) frequency that were extracted from every channel of wavelet decomposition using speech analysis tool, PRAAT. One level wavelet decomposition is conducted using Haar mother wavelet. Two types of statistical analysis, 2-way ANOVA method and Tukey HSD, were applied to determine the best channel of wavelet decomposition. In this process, we analized the effects of fundamental and formant frequency. Based on the experiment, we found that approximation channel (A) is better than details channel (D) as features for application in speech recognition, while detail channel (D) is better for speaker recognition. This result was evidenced by the formation of voice group using Tukey HSD, where in 1st level of decomposition details channel (11D) generates 15 voice group and approximation channel (10A) generates 14 voice group. On the other hand, in 2nd level of decomposition details channel (21D) generates 16 voice group while approximation channel (20A) generates 13 voice group. The higher voice group result shows that details channel (D) contain higher speaker information, contrary from approximation channel (A).