Sixth European Conference on Speech Communication and Technology
In this paper we propose an algorithm for efficient compression of feature extracted parameters used in speech recognition. The algorithm provides a compression ratio of roughly 1:10 and causes negligible or no loss in recognition performance. It is also shown to be robust against enviromental noise. Combined with an appropriate framing structure, a complete system is obtained, which can be used for implementing speech recognition applications e.g. in a cellular mobile environment. The system achieves a gross bitrate as low as 4200 bps.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. Kiss, Imre / Kapanen, Pekka (1999): "Robust feature vector compression algorithm for distributed speech recognition", In EUROSPEECH'99, 2183-2186.