Time and frequency domain algorithms for speech coding
The promise of digital hardware economies (due to recent advances in VLSI technology), has focussed much attention on more complex and sophisticated speech coding algorithms which offer improved quality at relatively low bit rates. This thesis describes the results (obtained from computer simulations) of research into various efficient (time and frequency domain) speech encoders operating at a transmission bit rate of 16 Kbps. In the time domain, Adaptive Differential Pulse Code Modulation (ADPCM) systems employing both forward and backward adaptive prediction were examined. A number of algorithms were proposed and evaluated, including several variants of the Stochastic Approximation Predictor (SAP). A Backward Block Adaptive (BBA) predictor was also developed and found to outperform the conventional stochastic methods, even though its complexity in terms of signal processing requirements is lower. A simplified Adaptive Predictive Coder (APC) employing a single tap pitch predictor considered next provided a slight improvement in performance over ADPCM, but with rather greater complexity. The ultimate test of any speech coding system is the perceptual performance of the received speech. Recent research has indicated that this may be enhanced by suitable control of the noise spectrum according to the theory of auditory masking. Various noise shaping ADPCM configurations were examined, and it was demonstrated that a proposed pre-/post-filtering arrangement which exploits advantageously the predictor-quantizer interaction, leads to the best subjective performance in both forward and backward prediction systems. Adaptive quantization is instrumental to the performance of ADPCM systems. Both the forward adaptive quantizer (AQF) and the backward oneword memory adaptation (AQJ) were examined. In addition, a novel method of decreasing quantization noise in ADPCM-AQJ coders, which involves the application of correction to the decoded speech samples, provided reduced output noise across the spectrum, with considerable high frequency noise suppression. More powerful (and inevitably more complex) frequency domain speech coders such as the Adaptive Transform Coder (ATC) and the Sub-band Coder (SBC) offer good quality speech at 16 Kbps. To reduce complexity and coding delay, whilst retaining the advantage of sub-band coding, a novel transform based split-band coder (TSBC) was developed and found to compare closely in performance with the SBC. To prevent the heavy side information requirement associated with a large number of bands in split-band coding schemes from impairing coding accuracy, without forgoing the efficiency provided by adaptive bit allocation, a method employing AQJs to code the sub-band signals together with vector quantization of the bit allocation patterns was also proposed. Finally, 'pipeline' methods of bit allocation and step size estimation (using the Fast Fourier Transform (FFT) on the input signal) were examined. Such methods, although less accurate, are nevertheless useful in limiting coding delay associated with SRC schemes employing Quadrature Mirror Filters (QMF).