Font Size: a A A

Research On 8kbit/s Low Delay Speech Coding Algorithm

Posted on:2008-12-07Degree:MasterType:Thesis
Country:ChinaCandidate:J F ShenFull Text:PDF
GTID:2178360242958981Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
The LD-CELP algorithm with low coding rate and low complexity has very important meaning in the field of communication. G.728 was the only 16kbit/s ITU Recommendation in the low delay speech coding algorithm at present. This research made improvement in G.728 algorithm aiming at reducing coding rate and presented three 8kbit/s speech coding algorithms whose delay are 2.5ms. The experiments showed the improved algorithms in the aspect of subjective and objective quality were close to that of G.728.Speech coding algorithm with low delay must adopt real-time pitch detection, but so far pitch detection has been long delay algorithm based on frame length more than 20ms and not meet the requirement for real-time. So G.728 can't adopt pitch detection. Real-time pitch detection is continuously detecting sampling signal of each frame, when a pitch pulse exists in sampling points of a frame, it is immediately detected and pitch is computed. This research designed real-time pitch detect algorithm whose frame length is 2.5ms using wavelet coefficients. Extremum of the 4th order wavelet coefficients is strongly related with sharp variation points of speech signal based on the multi-resolution analysis characters of wavelet. Based on which, real-time pitch detect algorithm based on wavelet transform was designed. This algorithm detected pitch through frame, combining waveform of wavelet domain with time domain, adopting adaptive criterion and multi-feature parameter to extract the maximum of wavelet coefficient. The algorithm could exactly detect pitch, start points and end points and the numbers of pitch based on extracting peak value of signal, which was the base of 8kbit/s low delay coding algorithm of backward pitch predictor.Adaptive codebook structure was introduced in G.728, which was composed of latest historical excited information, and normalized fixed codebook was adopted for the proposed algorithm. When fixed codebook was trained, it was optimally designed according to the frequency for use and the optimal fixed codebook was obtained by iteratively training. Then, fixed quantization was used in adaptive codebook and adaptive quantization was used in fixed codebook.Three schemes of 8kbit/s coding whose delay is 2.5ms were compared in this paper. The first scheme adopted whole search mode, that is searching the whole combination of 128 adaptive codeword, 8 gain and 128 fixed codeword, 8 gain, then the optimal index of codeword and gain were obtained; the second scheme adopted different search mode for adaptive codebook in continuous two frames: the search of the even frame was the same as the first scheme; the search of the odd frame was only done in the small area according to the index gotten in the even frame; the third scheme added backward pitch predictor to the first scheme. 64 codeword was searched according to pitch predict value in adaptive codebook, and 1 bit saved was used in extending the size of fixed codebook. Consequently, the subjective and objective quality of coding algorithm was improved.
Keywords/Search Tags:wavelet transform, pitch detection, adaptive codebook, gain quantization, SNR, PESQ
PDF Full Text Request
Related items