Non-linear Prediction of Speech Signal Using Artificial Neural Nets

Speech technology is one of the key technical issues involved in Information Technology as it constitutes an important aspect of Human Computer Interaction. Prediction of speech signal has applications in speech technology, especially in coding. Conventionally linear prediction is used. However, non-linear phenomena exist in speech production. Therefore, considering this non-linearity should lead to lower signal dynamics during coding with a consequent reduction in bit-rate and the needed bandwidth. This is studied in this paper using Feed Forward and Recurrent Neural Nets. It is shown through different evaluation schemes that the speech non-linearity is negligible and that non-linear speech prediction does not lead to an appreciable further reduction in the residual signal to be coded.