Signal modeling enhancements for automatic speech recognition

Obtaining a compact, information-rich representation of the speech signal is an important first step in ASR. A large majority of ASR systems use some form of cepstral coefficients for this purpose. Computation of these cepstral coefficients typically includes several of the following steps: (1) Highfrequency preemphasis, using an FIR filter of the form y(k) = x(k) ax(k-1), with a taking values around 0.95; (2) partition of the signal into analysis frames of 20 to 30 ms, spaced 5 to 10 ms apart; (3) computation of ten to forty cepstral coefficients using a cosine transform of the logarithm of the output of a 40-channel triangular filter bank, which is designed to approximate a Bark frequency scale; and (4) Feature vectors are assembled from the instantaneous cepstral values, augmented with some form of dynamic information, e.g. delta-cepstra. This paper describes several enhancements to this procedure. We show that significant improvements in recognition accuracy can be achieved by modifications in all of these steps, particularly for speech corrupted by noise. In particular, we show that 1. The first order high-frequency pre-emphasis should be replaced by a second order preemphasis of the form: