Research On Joint Source-Channel Decoding And Applications To Speech Transmission

With the rapid development of wireless communication and network communication, there has been an increasing demand for multimedia communication, and the technologies related to speech, audio and video has been drawn more attentions in the research filed of communication. Shannon's separate theorem is the design foundation of multimedia communication system, in which source coding and channel coding are treated as two independent subjects. Source coding aims to reduce as much redundancy as possible so that the source can be represented with fewer bits while the goal of channel coding is to make the transmitted signals robust to channel noise distortions and ensures the reliable transmission. However, the separate theorem is valid under the assumption of infinite-length codeword (resulting in infinite delay and system complexity) and point-to-point communication. This lead to the research of joint source-channel coding which jointly designs the source encoder (decoder) with channel encoder (decoder) to improve the system performance.This thesis investigates the joint source-channel decoding algorithm with application to speech transmission. The thesis proposes the joint source-channel decoding based on iterative structure and parameter estimation. According to priori information provided by time correlation of source parameter and channel output, channel decoder executes in the iterative process. Source decoder calculates parameter's posteriori probability based on bit reliability from iterative channel decoder and residual redundancy of source parameter. Then source parameters are estimated by MMSE (Minimum Mean Square Error) or MAP (Maximum a Posteriori) criteria. Also the posteriori probability is feedback as priori information to iterative channel on the next time. In order to avoid the transform of log-likelihood and probability, simplified joint source-channel decoding algorithm is also proposed to reduce the complexity. The simulation result shows that joint decoding obviously improves the SNR (Signal to Noise Ratio) of decoded parameter so that improves the system performance.The iterative channel decoder is serially concatenated because the convolutional encoder is of the non-systematic form in this thesis. Iterative channel decoder is composed of soft-in/soft-out Viterbi algorithm (SOVA) and bit estimation model. Based on EXIT (Extrinsic information transfer) charts of SOVA and bit estimation model, convergence behavior of iterative channel decoding is analyzed and the performance of joint decoding is determined by channel SNR and source redundancy. Also, a small number of iteration will greatly improve the performance. In general, the number is set to 2.Utilized tools of statistical analysis, the performance of source decoder based on MMSE estimation are analyzed. Under the assumption of jointly Gaussian distribution of source parameter and channel output, SNR of received parameter is deduced and the factors which determine the improvement of system performance are given. The performance improvement of source decoder on MMSE estimation is related to channel condition and the 1st auto-correlation of source parameter. Also, simulation proves the validity of analysis process.The residual redundancy inherent in the GSM EFR codec parameters is estimated. The redundancy due to the non-uniform distribution and inter-frame correlation is utilized as priori information both in channel decoder and source decoder. In order to exploit the redundancy as much as possible and avoid the priori probability of source parameter being used twice, the modification of joint source-channel decoding algorithm mentioned above is proposed and is applied to GSM EFR speech transmissionsystem. Also the estimation criteria (MMSE or MAP) are selected according to the property of speech codec parameters. The number of iteration channel decoding for GSM EFR is set to 1 based on the EXIT analysis result of iterative channel decoder. The simulation shows that the quality of received speech provided by joint source channel decoding is improved obviously compared with that of separated source channel decoding.