Perceived speech quality of Linear Prediction (LP) analyzed speech compressed by means of interpolation and quantization of pseudoformant parameters

Abstract

It is possible to produce high quality LP speech with a low bit rate by interpolating pseudoformants and bandwidths derived from the LP coefficients. To the extent that pseudoformants are like the true formants of speech, the importance of each can be inferred from classical acoustical studies. A listening experiment was performed which compared the perceived quality of speech interpolated using several assignments of importance to the pseudoformant parameters: an acoustically reasonable assignment, a uniform assignment, and a number of assignments intermediate between the two. Several levels of total distortion were used. The test shows that if an acoustically reasonable assignment of importance to each pseudoformant is given, the quality of the speech is significantly higher than if equal importance is given to each parameter. By using the acoustically reasonable allocation, very good speech quality can be maintained at a bit rate of 3200 bits/s while if the uniform allocation is used, more than 4200 bits are needed to attain the same quality. [Work supported by NSF.]