| 摘要 |
In a method for compressing feature extracted parameters of a speech waveform, at least a first set of selected ones of the feature extracted parameters are transform coded to provide a plurality of transform coefficients, and at least one of the transform coefficients is differential pulse code modulated (DPCM). Mel-frequency cepstral coefficients <I>c<SB>1</SB></I> to <I>c<SB>12</SB></I> are buffered (36, Fig.5a) and then transformed (38) using a discrete cosine transform (DCT) to form DCT coefficients DCT<I><SB>x,1...8</SB></I>. The first DCT coefficients are DPCM encoded (Fig.5b) with periodic pulse code modulation (PCM) updates to ensure synchronisation between a transmitter and a receiver.. PCM is applied to the other DCT coefficients, these coefficients being quantized using probability density function (pdf)-optimised quantizers. The zero order cepstral coefficient <I>c<SB>0</SB></I> and the logarithmic energy <I>logE</I> are DPCM coded with periodic updates similarly to the first DCT coefficients <I>DCT<SB>x,1</SB></I> . The memory requirement of the compression method is such that a front-end (FE) for feature extraction and a compressor can be incorporated in a mobile telephone handset, a decompressor and a back-end (BE) for speech recognition being located at a central place in a cellular network. |
