Third International Conference on Spoken Language Processing (ICSLP 94)
The work presented in this paper was devoted to the influence of telephone line distortions on speech recognition over the telephone network. It aimed at improving recognition performance by reducing telephone line effects in speech features to a minimum. Taking some measurements as a basis, we first show that for a given call, the telephone channel may be represented by a linear filter slowly varying with time. We then propose a theoretical framework showing that low frequencies of cepstral trajectories may help identify telephone channel effects. Identifying channel effects by means of low frequencies of cepstral trajectories confirms the interest of conventional techniques, such as high-pass filtering (RASTA) or cepstral subtraction, to increase the robustness of speech recognition over the telephone network. In this work, cepstral subtraction and high-pass filtering of the cepstral trajectories are experimented on a field database (recorded from a voice server in real use). All the experiments show that both techniques significantly improve recognition performance.
Bibliographic reference. Mokbel, C. / Paches-Leal, R. / Jouvet, D. / Monné, J. (1994): "Compensation of telephone line effects for robust speech recognition", In ICSLP-1994, 987-990.