A Sub-Band Coder Designed for Combined Source and Channel Coding
Richard Vandervoort Cox, Joachim Hagenauer, Nambi Seshadri · 1988
Subband coding has been shown to provide high quahty speech at rates of 16 kb/s and above. In dus paper we descnbe how one pmicular sub-band coder has been mded to give good quaky speech over a noisy channel. The coder is a dynmc bit allocation sub-band coder of the type 6rst proposed by Ramstad m ctured m such a way that the relative lmportance of all bits is byproduct of the dynmc bit allocahon. It is shown that there is rror sensitivity of four orders of maptude between the most and the least important bits of the ht stream on an average A very flemble unequal error protection scheme is used to match the error sensitivity of the sub-band coder bitst” llus is accomphshed usmg the concept of rate compatible punctured convoluhond coding. These treks codes are decoded by means of the Viteh algontbm usmg soft declslons on the nasy recaved slgnal. If necessary, the decoder can also mclude dormation about the time varymg channel characteristics in a rather straightforward way to lmprove performance. All ths is accomplished wthout any change in the encoder and the decoder stmctwe during changes in the level of error protection. The resulhng coder is more robust to channel errors than one m wbch all the error protechon is spread evenly among all bits or one m whch all error protechon is used for just a small number of important hts. The resultmg coder gives robust performance over a simulated noisy channel. Its nommal ht rate is 16 kb/s. wth 12 kb/s asslgned for speech codmg and 4 kb/s asslgned for enur correchon. The enhre encoder and decoder &T DSP-32 digpal signal processor. speech is a relatively mature form of waveform speech is first subdwided into a number of subndividually encoded. The underlying pnnciple for the coder is that the bit docahon can be weighted so that those subbands with the most impomt Informahon get the most bits. The imtial sub-band coders used fixed bit allocations based on the average spectrum of speech. Typical of hs generation of coders was one by Crochere [l]. In 1982 Ramstad introduced the idea of dynamically changmg the bit allocation [2]. ms idea was to qUanhZe and transmit the rms value of each of the bands for a frame of speech. Based on the quanhzed values, the remaning bits could be allocated among the sub-bands m an optimal fashion. Most recently, Honda & ltakura [3] and Soong, Cox & Jayant [4] have proposed dynanucally allocahng bits m both time and fkquency. Theu work produces very hgh qualtty speech. The complexlty of these algorithms is very hgh, however. In the work reported here, we have considered a coder based on the ideas proposed by Ramstad. The speech is divided into six 500 Hz wide sub-bands and mto 16 msec frames. Once every frame, the rms value of the speech m each band is quanbed and transmitted. Ths “side informahon” uses 2 kb/s of the bit rate. Based on the side informahon, an adaptive bit allocahon can be performed. Assuming the side informahon is correctly received, the same bit docahon can be denved from the side mformahon at both the transmitter and receiver. Nonlinear PCM is used to quanhze the sub-band samples. No m-band predxtion is used The side informahon can also be used to set the quanhzer step sue for each band. The man information, consisting of the quanhzed values of the samples of each sub-band, make up the majonty of the mformahon on the channel. Figure 1 s a block bagram of the transmutter pomon of hs coder. We shall refer to hs coder as dynamic bit allOCahOn SBC or D-SBC and the earher coder of Crochere [l] as ked bit allocahon SBC BC.