Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Compensating for Additive-Noise in Automatic Speech Recognition

Solomon Lerner, Baruch Mazor

GTE Laboratories Inc., Waltham, MA, USA

Mismatch in the quality and characteristics of speech to be recognized and training speech significantly affect the performance of automatic speech recognition (ASR) systems. Since our applications are over the public switched telephone network (PSTN), we have been investigating the impact of the network on the speech features and recognition and have explored several compensation strategies. To this end, we developed [1] a normalization method to compensate for the effects of spectral shaping (linear filtering) in DTW recognizers. This compensation method, essentially a cepstral subtraction operation, improves recognition performance significantly. In this paper, we extend the work in [1] to address the effects of additive-noise and to modify the cepstral subtraction method for HMM recognizers.

Full Paper

Bibliographic reference.  Lerner, Solomon / Mazor, Baruch (1992): "Compensating for additive-noise in automatic speech recognition", In ICSLP-1992, 723-726.