Research On Very Low Bit Rate Speech Coding

In the study of speech coding in present,the Mixed Excitation Linear Prediction(MELP) is a kind of relatively good method. The MELP model as defined in MIL-STD-3005 is based on the traditional LPC-10e parametric model, but also includes five additional features. These are mixed excitation, aperiodic pulses, pulse dispersion, adaptive spectral enhancement, and fourier magnitudes scaling of the voiced excitation. It can simulate more natural speech characteristics and can achieve high synthesis speech quality. This paper presents very low bit rate speech coders based on the MELP analysis algorithm. The works are as follows.One fundamental issue for the LBG algorithm is the problem of the local minimum, where there is no guarantee that the codewords after convergence can provide the best possible solution, or global minimum. As a result, simulated annealing(SA) algorithm is used to perform the vector quantification codebook. Compared with LBG, the average distortion decreases.Theoretics analysis and experiment test by adding noise, the quantization characteristic of LSF is validated to be excellent. A efficient LSF quantization methods that is predictive switched multi-stage vector quantization(PSMSVQ) is proposed. Compared with split vector quantization (SVQ) and multi-stage vector quantization(MSVQ), its quantization performance is improved at the expense of higher memory requirements for storing the codebooks.As parameters extraction of bandpass voicing strengths is an unprecise method, a more strict sub-band rule is used so that hearing sense is taken into account. As the preprocessing highpass filter and adaptive spectral enhancement filter can produce negative effection, the compensating method is applied and improve the low-pitch male voice.Based on the model of MELP, in order to exploite interframe and intraframe correlation present among the parameters, multi-frame joint, the high efficient quantization differently according superframe mode, parameter spectrum comparability interpolate and frame length enlargement techniques are promoted in this paper.Two kinds of very low bit speech coders at 800bps and 600kbps are designed and implemented in this subject.Subjective listening tests show that the speech quality of both proposed vocoders is better than that of traditional LPC 2.4kbps algorithm, lower than that of standard 2.4kbps MELP in some sort.The synthesized speech of 800bps coder achieves high intelligibility, clearness, and relative naturalness,600bps coder also get high intelligibility.