Work

Variable rate speech encoding using linear predictive methods

Year: 1980
Type: dissertation
Abstract: In analysis-synthesis systems like Linear Predictive Coding (LPC), the speech signal is modeled as the output of a time-varying digital linear filter representing the vocal tract, where the filter input is random noise or a sequence of impulses. In LPC, this filter is described by a set of parameters, called PARCOR coefficients, which are then coded to reduce (compress) the bit rate necessary to code the speech signal. An equivalent set of parameters are the area functions, through which the vocal tract is modeled as a concatenation of cylindrical tubes. This thesis addresses the problem of further reducing the speech bit rate by manipulation of the LPC parameters without lowering the quality of the resynthesized speech signal. In this study, the vocal tract is viewed both statically and dynamically. In the static description, certain sections of the vocal tract model are unified so that the vocal tract is represented by fewer tubes. The selection of these sections is based on the minimization of an objective speech distortion measure. The new reduced model is tested for both its effectiveness in modeling the vocal tract and its speech compression capabilities. The dynamic description is connected with the time varying nature of the vocal tract filter. The speech signal is divided into quasi-stationary frames where the filter parameters remain constant within each frame. For certain phonemes (e.g. vowels), the filter parameters often will not change greatly from one frame to the next and when successive frames are encoded, some or all of the parameters can retain their previous values. Hence, it is possible to transmit some, all or none of the new LPC parameters. This choice is not made immediately; instead, a number of consecutive frames is considered and all the possible decision sequences are examined. The final choice of the optimum decision sequence is based on the minimization of a cost function which penalizes both for the number of bits transmitted per frame and for the distortion introduced in the speech signal. The examination of the possible decision sequences is equivalent to a decision tree search which is most efficiently done through dynamic programming. The quality of the resynthesized speech was judged by listeners, who ranked the different systems in a user preference test. The reduced static model was judged to be a very effective representation of the vocal tract, but its speech compression capabilities were limited because of the necessary additional overhead information. On the other hand, the dynamic model achieved a high compression of the data rate: 1200 bits per second speech at the output of the present systems was judged to be of equal quality to 2400 bits per second speech generated by only quantizing the LPC parameters. These results show that the dynamic model is very effective for speech compression in voice response applications. It has less importance for real time applications because of the required long analysis time.
Language: English
Open Access status: green
OpenAlex ID: W2499712276
Cites:
Cited by: 2
Related to: 10