A stochastic model excitation source for linear prediction speech analysis-synthesis

The aim of this paper is to simplify the new speech analysis-synthesis based on stochastic model of speech production proposed by Atal et. al.. Both the coder and the decoder have the same noise inventory. The prediction error obtained after both short and long term predictin of input speech is successively compaired with the noise inventory adjusting the scale factor. The code number which indicates the optimum noise sequence is sent to the decoder together with side informations concerning short and long term prediction. In the decoder, noise sequence specified by the code number received are sequentially passed through two filters: the first one with long term prediction and the second one with short term prediction. The experimental results show that the synthesized speech is good in quality. The transmission rate is estimated as low as 8 kbps. The method for generating an efficient noise inventory is also presented in this paper.