New linear predictive methods for digital speech processing

Abstract

Speech processing is needed whenever speech is to be compressed, synthesised or recognised by the means of electrical equipment. Different types of phones, multimedia equipment and interfaces to various electronic devices, all require digital speech processing. As an example, a GSM phone applies speech processing in its RPE-LTP encoder/decoder (ETSI, 1997). In this coder, 20 ms of speech is first analysed in the short-term prediction (STP) part, and second in the long-term prediction (LTP) part. Finally, speech compression is achieved in the RPE encoding part, where only 1/3 of the encoded samples are selected to be transmitted.

This thesis presents modifications for one of the most widely applied techniques in digital speech processing, namely linear prediction (LP). During recent decades linear prediction has played an important role in telecommunications and other areas related to speech compression and recognition. In linear prediction sample s(n) is predicted from its p previous samples by forming a linear combination of the p previous samples and by minimising the prediction error. This procedure in the time domain corresponds to modelling the spectral envelope of the speech spectrum in the frequency domain. The accuracy of the spectral envelope to the speech spectrum is strongly dependent on the order of the resulting all-pole filter. This, in turn, is usually related to the number of parameters required to define the model, and hence to be transmitted.

Our study presents new predictive methods, which are modified from conventional linear prediction by taking the previous samples for linear combination differently. This algorithmic development aims at new all-pole techniques, which could present speech spectra with fewer parameters.