Deconvolution of telephone line effects for speech recognition

Abstract This paper presents a new approach to equalize the telephone line effects in the transmitted signal aiming at improving the performance of speech recognition systems. This new approach implements a blind equalization scheme where an adaptive filter, using some known statistics about the signals, deconvolves the channel from the transmitted signal. Measurements carried out on actual telephone data confirm that telephone lines introduce disturbing convolved components in speech signals. Line effects are almost constant for a given call but vary with the calls. The proposed adaptive filtering of the telephone line effects is compared to two conventional techniques, namely subtraction of the long-term cepstrum and highpass filtering of cepstral trajectories. Recognition experiments are carried out on several telephone databases in a speaker-independent mode. The results show that reducing the channel effects significantly improves the recognition performance. Regarding the obtained error rates, the proposed adaptive filter yields better performance than the conventional highpass filters. However, this adaptive filtering is not as good as the off-line cepstral subtraction technique where the long-term cepstrum is estimated on several recordings of a call. Experiments were also conducted to measure the amount of speech data necessary to obtain a reliable estimate of channel effects. Averaging cepstra vectors on a few seconds of speech produces a reliable estimate of the constant convolved perturbation.