Intelligibility Performance of Narrowband Linear Predictive Vocoders in the Presence of Bit Errors.

摘要

Diagnostic speech intelligibility tests were evaluated to assess vulnerability of two different 2400 bit-per-second linear predictive vocoder algorithms to random bit errors imposed on the data stream. Listening tests with crews of eight subjects yielded diagnostic intelligibility scores at zero, 1%, 3% and 5% bit error rates. These data were analyzed to establish linear regression models relating intelligibility performance and bit error rate. Piecewise-linear prediction coding (PLPC) was confirmed to have a small but significant advantage through being less vulnerable to bit errors than conventional linear prediction coding (LPC), an advantage that had been hypothesized from the inherent redundancy that is added by transmitting separate LPC coefficients for low-frequency and high-frequency speech bands. A small but consistent improvement in intelligibility was also found for the error-free case, believed to result from improved spectrum modeling that is a consequence of the piecewise approach. Significant differences in susceptibilities to bit errors were found among individual intelligibility scores for speakers as well as for intelligibility features. Tables for predicting average intelligibility performance, and confidence limits, were constructed from the regression models. The findings provide guidance for further research towards the goal of minimizing susceptibility of narrowband LPC vocoders to jamming and interference. They also highlight a need for further studies to obtain better understanding of causes of the typical large dispersion in intelligibility scores for individual speakers, obtained in these and many other tests.