Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Blind Equalization Using Adaptive Filtering for Improving Speech Recognition Over Telephone

C. Mokbel, D. Jouvet, J. Monné

France Telecom - CNET/LAA/TSS/RCP, Technopole Anticipa, Lannion, France

It is well known that the telephone channel introduces a convolved disturbing component in the observed signal which slowly varies with time. Several techniques have been proposed in the literature to reduce the telephone line effects, namely cepstral subtraction and highpass filtering of cepstral trajectories. In a previous work [8] we introduced a theoretical framework to clarify the link between the channel effects and the low frequencies in the cepstral trajectories and to interpret the encouraging results obtained by cepstral subtraction (an off-line approach) and highpass filtering of the cepstral trajectories. This paper proposes a new approach to reduce the telephone line effects. This approach implements an adaptive filter in a blind equalization scheme. On the basis of recognition results obtained on different speaker independent telephone databases, the blind equalization seems to outperform the classical on-line technique.

Full Paper

Bibliographic reference.  Mokbel, C. / Jouvet, D. / Monné, J. (1995): "Blind equalization using adaptive filtering for improving speech recognition over telephone", In EUROSPEECH-1995, 1987-1990.