Next: ADDING ADDITIONAL FEATURES Up: A COMPARISON OF Previous: FILTERBANK AND THE

LPC BASED TECHNIQUES

Linear Predictive Coding (LPC) is a very popular front end for speech recognition systems because of the low computational cost compared with FFT or filterbank methods. All the preprocessors considered in this section will use 16th order LPC computed using the autocorrelation method. acf is just the autocorrelation values, lpf is the resulting linear predictor filter coefficients, lpa is the log area ratios of the equivalent lossless tube, cep are the cepstral coefficients derived from LPC and l20 is a the cube rooted 20 sample bark scale spectrum derived from the LPC filter. In addition, the Smoothed Group delay model of Singer, Umezalia and Itakura was included because of its good quantisation properties [6].

It is interesting to note that although all these representations are derived from the same set of autocorrelation coefficients, there is a significant range in performance. lpf and acf which are linearly related to the power spectrum ( pc1), show poorer performance even though this mapping can be incorporated into the first layer of weights at no additional cost. l20 shows slightly worse recognition than p20, presumably because of the smoothing imposed on the power spectrum by the LPC representation.

  
Table 7: LPC based preprocessors