Research On Ultra-Low Bit Rate Speech Coding Algorithms

Ultra-low bit rate speech coding techniques have been important research topicsfor many universities and research institutes. They are widely used in short-wavecommunications and underwater acoustic communications. In addition, in militarysecure communications and satellite communications which channels are expensive,ultra-low bit rate speech coding techniques have also been widely used.In the ultra-low bit rate speech coding, the number of bits assigned to eachfeature parameter is extremely limited, which brought great difficulties to thequantization of parameters. In order to solve the problem, an algorithm usingsecond-order Hidden Markov Model (HMM2) to recover the voiced/unvoicedparameters has been proposed in this paper. The algorithm uses the normalized energyand linear prediction coding (LPC) coefficients to estimate the full-band V/Uclassification and the sub-band BPVC value. The algorithm can be implemented inthe decoder, saving the bits originally used by V/U parameters and reducing the bitrate of speech coding.Multi-frame coding technique is often applied in low bit rate speech coding dueto its excellent quantization performance. But with the lower encoding rate, thenumber of joint frames corresponding increases, which makes the time complexityand space complexity of algorithm also increase. To solve the problem, this paperpresents the Important Frame Quantization Algorithm for Line Spectral Frequency(LSF) parameters which based on statistics; the algorithm does not affect the qualityof synthesized speech, greatly reducing the time complexity and codebook storage.Based on the above techniques and other mature techniques, parameter-related150bps vocoder is designed and implemented. The codebook storage of theparameter-related150bps vocoder is86K words. Tests show the objective meanopinion score (MOS) of the150bps vocoder is2.41, and more than82%of the singlewords synthesized by the vocoder could be recognized correctly. The performancealready exceeds the demand for Eleventh Five-Year Plan.Finally, an Automatic Gain Control (AGC) module is designed and implemented for the practical application of the vocoder. The AGC modulecan adaptively control the input signal to guarantee its amplitude in a modestrange. In addition, in order to increase the robustness of the2.4Kbps vocoder,an algorithm using the intra-frame relevancies of line spectral frequency(LSF) parameters to detect the error of LSF parameters and to recover themwhen they have been detected has been proposed. Tests show that thealgorithm can improve the robustness of vocoder greatly.