Low bit-rate vector excitation coding of phonetically classified speech
Shihua Wang · 1992
As the demand grows in the wireless information networks for precious communication channel bandwidth, so does the need for high quality speech coding at rates of 4 kb/s or below. Many promising low-rate-coding schemes have been proposed in recent years to bridge the gap between toll quality waveform coding at rates above 16 kb/s and poor quality vocoding at rates around 2.4 kb/s. Specifically, analysis-by-synthesis predictive coding has produced good speech quality at rates down to 4.8 kb/s, but now the effort is focused on the 4.8-2.4 kb/s range. While signal processing techniques for data compression have matured, (especially vector quantization which serves as the cornerstone for low-rate coding algorithms), they cannot by themselves bring about a further reduction of bit rate. Rather, we must also understand better the human articulatory and auditory systems as part of the entire speech communication link, instead of merely striving to faithfully reproduce the acoustic waveform of speech. The main purpose of this dissertation is to integrate knowledge of speech production and perception into analysis-by-synthesis coding algorithm and thereby reduce the bit rate below 4 kb/s without sacrificing quality. As a first step, we examined the microstructure of coding distortions in this type of coders at 4.8 kb/s. Next we introduced phonetic classification as a pre-processing to coding, in an attempt to adapt coder behavior to phonetic content. Four phonetic categories were identified and were coded with different coding strategies. We showed that phonetic classification permits the reduction of the data rate to 3.4 kb/s with a minimal change in quality. We also address the important issue of speech quality assessment, by proposing a new perceptually-motivated objective measure for evaluating speech distortions. In a test speech database including 7 coders with rates ranging from 2.4 to 64 kb/s, the predictions of the measure correlated well (r = 0.96-0.98) with MOS scores.