SPEECH PARAMETER COMPRESSION

摘要

In distributed speech recognition feature extracted parameters need to be transmitted over the cellular network to the speech recognition engine. The existing networks, such as GSM or TETRA, however, cannot provide the necessary bandwidth on their data channels. The compression of feature extracted parameters is necessary. Special problems have to be addressed during the design of the compression algorithm. Compression must have no effect on recognition accuracy. The algorithm must be robust against environmental variations and noise. Its computational complexity and memory requirement must be moderate. This invention describes a compression scheme which fulfils the above requirements, and makes distributed speech recognition applicable in cellular environments. The compression scheme disclosed consists in a method for compressing feature extracted parameters of a speech waveform for transmission and/or storage comprising transform coding a first set of selected ones of said feature extracted parameters to provide a plurality of transform coefficients and differential pulse code modulating at least one of said plurality of transform coefficients. The method further comprises differential pulse code modulating a second set of selected ones of said feature extracted parameters.