Sixth European Conference on Speech Communication and Technology
In real-world speech and speaker recognition systems, data is often recorded over commercial telephone lines. Consequently, differing transmission channels cause mismatch between training and testing conditions resulting in significant performance loss. This paper presents a new technique that uses complex spectral averaging to estimate the channel accurately. The estimated channel is used as an inverse filter for normalization. This technique being speech-in speech-out, can be used as the preprocessing stage in any automatic speech processing system. A refinement process is also presented that further improves the channel estimate. The combined technique is evaluated on a speaker verification task where the training and testing data were convolved with different telephone channels. The new technique provides excellent channel estimates and nearly restores performance back to that of clean conditions.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Balchandran, Rajesh / Ramanujam, Vidhya / Mammone, Richard J. (1999): "Channel estimation and normalization by coherent spectral averaging for robust speaker verification", In EUROSPEECH'99, 755-758.