A study on a reduction of the transmission bit rate by U/V decision using LSP in the CELP vocoder

In a typical CELP vocoder, the unvoiced speech is processed like the voiced speech. Hence the speech quality is degraded. Also, the degree of the speech quality degradation is excessive depends on the performance of the U/V classifier. In this paper, we propose a new U/V (Unvoiced/Voiced) decision algorithm with minimized error rate and preprocessing computation. This U/V classifier uses the LSP (Line Spectrum Pair) parameter, which is acquired from the spectrum analysis process in CELP vocoders. Applying this method to the 5.3 kbps ACELP, we can achieve a transmission bit rate reduction of 5.6% (280 bps) approximately. By applying the synthesis speech quality in the G.723.1 and the proposed method to the MOS (Mean Opinion Score) test the former gets average MOS 3.77 and the latter does 3.75