Sample Reduction and Subsequent Adaptive Interpolation of Speech Signals

Sample Reduction and Subsequent Adaptive Interpolation of Speech Signals By R. STEELE* and F. BENJAMIN 1, (Manuscript received December 15, 1982) In this paper we investigate the effect of rejecting every nth speech sample and replacing it by means of adaptive interpolation. The interpolation procedure attempts to minimize the mean square interpolation error by recomputing the autocorrelation function of the speech sequence every W samples. We describe three methods of computing the correlation function. An iterative procedure is evaluated for estimating the correlation function of a speech sequence whose every nth sample has been discarded. For speech bandlimited to 3.2 kHz, sampled at 8 kHz, and n = 4, W = 256, the gain in signal-to-noise ratio (s/n) achieved by adaptive interpolation compared to nearest neighbor average interpolation was 14 and 8 dB, depending on whether the correlation function was computed from the original speech, or by using the iterative procedure, respectively. The effect of varying n from 2 to 6 was also investigated. Finally, we applied the interpolation procedures to 8-bit |i-law pulse code modulation (PCM), n = 255, reducing 64 kb/s transmission to 48 kb/s by rejecting one PCM word in four. The recovered speech after interpolation had a s/n that approximated that of conventional 56 kb/s /x-law PCM speech. I. INTRODUCTION Sampling of speech signals is usually performed at a rate high enough to prevent objectionable aliasing. Thus a speech signal whose bandwidth extends from 0.3 to 3.3 kHz is typically sampled at 8 kHz.