Very Low Rate Low Delay Speech Coding Algorithm Research

This thesis mainly concerns about the research of LD-CELP speech coding algorithm in order to respond the new target proposed by ITU, an improved speech encoding algorithm reaches toll quality, with delay less than 5ms and coding rate less than 8Kb/s. The purpose toward all endeavors is to:1. Design a speech encoding algorithm which reaches toll quality with 20 samples per frame, delay less than 2.5ms and coding rate approximately 8Kb/s.2. Design a speech encoding algorithm which reaches communication quality with 40 samples per frame, delay less than 5ms and coding rate approximately 6Kb/s.To accomplish above the target, we researched in the following five aspects and have got effective progress.First, we have designed an effective method independent of quantizing signal noise rate (SNR) in evaluating the performance of gain filter. This provides a way to compare and evaluate different optimizing schemes before gain quantization, forms a way of independent research between gain predictor and gain quantizer. After the research and evaluation of different predictors, we found the recursive filter with finite memory gives the best performance and neural network filter, with low computing price, have better performance than G.728 Durbin recursive formula. We researched and compared the feature of fixed quantizing and adaptive quantizing. The difficulty in adaptive quantizing is the optimizing of adaptive step need mass computing. This thesis proposed a new method in optimizing multi-variable in N-coded complex target function based on clone method evolution algorithm in artificial immunization. This gives a new solution to solve the problem.Second, we developed a new backward real time speech pitch detection method with the help of wavelet analysis tools. The traditional pitch detection methods are forward long delay algorithm and all low delay speech encoding algorithm, including G.728, have no pitch detection. We delivered a backward real time speech pitch detection method with the capability of detecting the appearance of pitch and calculating the pitch cycles within 2.5ms, or in 20 samples. This makes it possible for the pitch analysis used in LD-CELP.Third, we employed the module of adaptive codebook searching in low delay speech encoding algorithm. The bit rate for each sample is less than 1 bit in low delay algorithm, so it is not easy to compact code rate while using adaptive code book. Our improving method is to give a raw locate of best adaptive code vector index using backward real time pitch detection algorithm first and performs the precise adaptive codebook searching based on it. This greatly reduces the code rate of adaptive code book, improves the precision of backward pitch detection and get the ideal effect when using in low delay speech encode algorithm.Fourth, we designed the new algorithm with delay of 2.5ms (20 samples per frame) and code rate of 8Kb/s. We expanded the frame length from 5 samples to 20 samples according to G.728 structure. It has been found by lots of experimentation when the code rate less than 8.8K/s, the speech quality deteriorated and the computing complexity is unacceptable in real time algorithm, so we have to configure a new algorithm structure. We researched and realized three algorithms with delay of 2.5ms and bit rate of 8Kb/s.Scheme 1: Both 10 bit for adaptive code book and fixed code book (3 bit for gain and 7bit for wave).Scheme 2: In an even frame searches adaptive code book and for sequential odd frame uses the searching result of preceding even frame, the saved bits is used to enlarge fixed code book size.Scheme 3: Adopts the scheme integrated the backward real time pitch detection into the adaptive code book searching method.By experimentations, all three methods have approached toll speech quality, with delay of 2.5ms and code rate of 8Kb/s.At last, it is discussed the method of further reducing the code rate. Different from other backward encoding algorithms, this research adopted algebraic code excited for fixed code book. We designed a new backward pitch detection method in combination of updating adaptive code book searching method, which reaches communication speech quality with delay of 40 samples (5ms) and bit rate of 6.2K/s.