Down-sampling speech representation in ASR

Features for automatic speech recognition (ASR) are typically sampled at about 100 Hz (10 ms analysis step). Recent experiments indicate that the most e cient components of the modulation spectrum of speech for ASR are up to about 16 Hz [1]. Consequently, RASTA processing attenuates modulation frequencies higher than 16 Hz and should in principle allow for a subsequent down-sampling of the features. It has been shown earlier that in a Gaussian mixture model based speaker recognition system(which uses single state HMM, thus not requiring any time alignments of the incoming speech) one could down-sample the speech representation after RASTA ltering without any signi cant loss of performance [2]. However since ASR uses Viterbi time alignment, reduced number of time samples due to down-sampling, although justi ed by Nyquist criteria after the low-pass ltering, could create problems. In this paper we experimentally show that the downsampling of features after RASTA ltering is feasible and could result in considerable computational or at least storage/transmission savings. 1 Temporal processing Speech contains many source of information such as information about the linguistic message, about the speaker of the message, and about the communication channel used for the recording and transmission of the speech signal. For a given task, it is helpful to retain relevant source of information in extracted features while suppressing the irrelevant ones. In ASR, the task is to decode the linguistic message This linguistic message is coded in the movements of the vocal tract. The speech signal re ects these movements. The rate of change of the non-linguistic components in speech often lies outside the typical rate of change of the vocal tract shape. The RASTA [3] and LDA [4] techniques take advantage of this fact and bandpass lter time trajectories of speech feature vectors. 1.1 Exploiting the bandpass property RASTA lters out the fast (and slow) changes of spectral components over time. Since fast changes (high modulation frequency components) are eliminated by RASTA ltering, the Nyquist criterion would suggest that the RASTA ltered features could be sampled at a sampling rate slower than that of original nonltered features(Fig 2).