Piecewise linear quantization of linear prediction coefficients
C.K. Un, Sheng-Jenn Yang · The Journal of the Acoustical Society of America · 1976
We present a new coefficient-coding method of linear predictive coding (LPC) of speech. It employs piecewise linear quantization and requires statistical properties of the LPC reflection coefficient {ki}. In coding the coefficients we first construct histograms of the distribution of each coefficient values obtained from analysis of many different speakers. Then, rather than making quantization levels uniform throughout the range of coefficient values, we divide the range of each coefficient or its transformed version into three regions and assign a different number of quantization levels to each region. Before the range of each coefficient is divided, the minimum value R0 and the maximum value R3 of each ki are set on the basis of experimental observations. Division of each range into three regions is made by specifying R1 and R2 according to the distribution of coefficient values and desired quantization accuracy in each region. Although the piecewise linear coding places its importance on the density of frequencies of the coefficient values, it does not neglect the importance of the spectral sensitivity aspect. In our informal subjective listening test it was observed that the quality of synthetic speech with the transmission rate of 2.4 kbits/sec coded by piecewise quantization was equivalent to the quality with the rate of 3 kbits/sec coded by linear quantization. The piecewise linear quantization scheme can also be used for the arc sine and log area ratio coding methods that normally employ linear quantization. In this case one can expect approximately the same amount of saving of transmission rate as for the nontransformed reflection coefficients.