Error correction scheme for telephone line transmission of RELP vocoder

Abstract

This paper presents a scheme for encoding the output bits of an 8500 bps residual excited linear predictive coder (RELPC), in which the quality of the synthesized speech is more sensitive to errors tn certatn btts.Fifty-two of the 272 bits in a 32 ms voice frame were identified as most critical and in need of protection. A Burst-trapptng code, using a (21,14) Single Error Correcting Double Error Detecting shortened Cyclic Hamming code as the basic block code was constructed. The code was selected with due consideration so computational load, delay sn decodtng and total bit rate requirement of less than 9.6 kbs. The paper also describes a Markov model for the decoding process and then determtnes the improvement in quality resulting from the use of the above scheme. In particular a bit error rate of 5 x l0 is reduced to 3 x 106 and a burst error rate of 5 x I o is reduced to 6.8 x I pS• Some of the effects of errors on auditory perception which appear counter intuitive are briefly dtscussed.Description of the RELP vocoder output.The RELPC vocoder used for this investigation [11, outputs 272 bits of vucoded information for each 32 ms interval (henceforth called a frame) of voice. The 272 hits consist of:1. 32 bits for the reflection coefficients of the vocal ttact filter.2. 20 hits for energy in the high frequency ( > 800 cycles) part of the residual excitation. 3. 200 bits for the residual excitation in 5 sub-bands between 250 and 875 cycles.4. 15 allocation bits. With 8 samples per sub-band, there are on the average 5 bits per sample per sub-band. However the 25 bits per sample are not equally distributed amongst the 5 sub-bands and the number of bits associated with each sub-band is encoded in an additional 3 bits, giving a total of 15 bits for the 5 sub-bands. 5. A 5 bit table-index for the quantization step sizes for the residual excitation of all sub-bands.The transmission of 272 bits in a interval of 32 ms, gives a transmission bit rate of 8.5 kbps. Since a 9.6 kbps modem is to be used, there are 35 remaining bits per frame which could be used as check bits for encoding the frame for crror correction. It is well known [2]-[5], that errors in different bits effect the quality and intelligibility of the synthesized speech to different degrees. Having performed some tests on the digitized speech files at our disposal, we obtained the following order, of decreasing sensitivity to the synthesized speech, among the bitt. a. The 32 bits of the 8 reflection coefficients.b. The 15 allocation bits. c. The 5 bits that give the quantizer step size of the residual signal.d. The 20 bits that give the high frequency energy. e. The 200 bits of residual excitation in the 5 sub-bands. It was therefore decided to protect only the 52 bits specified in a, b and c. Cnding considerations.A telephone line channel is a channel which exhibits both random and burst errors [6]. Codes that are suitable for such channels are given in [7,8]. According to these papers the Burst-trapping Codes give the best results. We designe the code to...