Research On Key Problems Of Vocoder System Based On MELP

Mixed excitation linear prediction (MELP) is a kind of low bit rate speechcoding algorithm. In order to improve the error resilience ability of vocoder systembased on MELP in transmission channel, forward error correction (FEC) coding isusually used. At the same time, noise degrades the performance of speech signalprocessing, so speech denoising is needed in front of vocoder. With the improvementof vector quantization for spectral line frequency (LSF) parameter of speech, FEC ofMELP vocoder was designed based on (2,1,2) convolutional code, and speechcontaining additive white noise was denoised by fast fixed-point algorithm (FastICA)of independent component analysis (ICA), as a result, the error resilience ability andcoding performance of MELP vocoder is improved.First of all, based on the design of optimum vector quantizer, the stage of vectorquantization for LSF coefficient was reduced from four to three, the bit rate of MELPis lower, while there is only a little loss of speech quality.Secondly, on the basis of Hamming code error correction in unvoiced speechframe, the most important23bits parameters in MELP stream of each frame wereprotected by (2,1,2) convolutional code, so FEC of MELP vocoder based onconvolutional code was designed, the bit rate increases from54bits/frame to81bits/frame, the error resilience ability of MELP vocoder in channel is also improved.In order to ensure the transmission reliability of MELP vocoder at a lower bit rate, thestage of vector quantization for LSF coefficient was reduced from four to three, indexbits of the fourth stage of code book were for convolutional coding, so the bit rate ofMELP vocoder based on convolutional code FEC decreases from81bits/frame to75bits/frame, however, the error resilience is not weaken.Finally, fast fixed-point algorithm of independent component analysis wasdesigned. As ICA requires that the number of observed signals must not be less thanthat of source signals which are independent of each other, in order to denoise singlespeech by ICA, one noisy speech was processed by spectral subtraction algorithm, sotwo noisy speech signals were constructed, and then FastICA algorithm processed thetwo observed signals, at last, denoised speech and noise as source signals wereseparated. The simulation results show that this method can improve speech to noiseratio.