TITLE:        A trellis-searched 16 kbit/sec speech coder with low delay
AUTHORS:      M. W. Marcellin and T. R. Fischer
CONFERENCE:   IEEE Workshop on Speech Coding for Telecommunications, 
	      Vancouver, B.C., Canada, September 1989
ABSTRACT:
Trellis coded quantization (TCQ) was recently introduced [1] as an 
effective scheme for encoding memoryless sources. TCQ is motivated by 
Ungerboeck's formulation of trellis coded modulation [2] and uses 
a structured codebook with an expanded set of quantization levels.  
Based on Ungerboeck's notion of set partitioning, the trellis structure 
then prunes the expanded number of quantization levels down to the 
desired encoding rate.  By employing a deterministic codebook and Viterbi 
encoding [3], a computationally simple encoding structure is achieved.  
 
TCQ has previously been incorporated into a DPCM structure for encoding 
speech at 2 bits per sample (16 kbps) [4]. The signal-to-noise ratio (SNR) 
and segmental SNR (SEGSNR) performance of this system is quite good. 
Specifically, for a variety of speakers and test sentences, values of SNR 
were obtained between 17.52 dB and 21.66 dB, while the SEGSNR ranged from 
19.13 dB to 21.90 dB.  The reconstructed speech obtained from this system 
was judged, in informal listening tests, to be of excellent communications 
quality.
  
In the work presented here, we extend the results reported in [4] by 
incorporating TCQ into a generalized predictive waveform coding structure 
with backward adaptive formant and pitch predictors. 
This structure is often referred to as a noise feedback coding 
structure [5]. Similar results for tree coding have been reported by 
Iyengar and Kabal [6]. The trellis encoding is accomplished by means of a 
modified Viterbi algorithm which allows flexibility in the choice of symbol 
release rule.  The effects of varying the encoding delay and the number of 
symbols released per trace-back on system performance and complexity are 
investigated.  In general, increasing the encoding delay increases the 
signal-to-noise ratio of the encoding, but this increase in 
performance saturates quickly.  For encoding sampled speech at 16 kbps, 
nearly all the gain achievable by the trellis search can be obtained with 
encoding delays of 5 msec or less.