Prediction of formant frequencies from linear combinations of filterbank and cepstral coefficients

The topic of this paper is formant frequency prediction using multiple linear regression. This technique provides robust formant frequency estimates but with limited precision. We apply this approach to predict formant frequencies of one male speaker comparing three different spectral representations. Mel scaled filterbanks are shown to perform slightly better than linear prediction based cepstral coefficients. The baseline approach uses one linear prediction model for each formant. This approach is extended by using multiple models for each formant. In this case, each model is trained on data from sub-bands of the formant frequency. This method is shown to be very useful for predicting F2, which has a large frequency span. The rms error of F2 can be reduced by up to 45% using multiple models on an independent test set. Moderate sizes of training data suffice to derive the linear prediction models which implies that predictors can be trained for a new speaker with a small formant labelling effort.