Recognition of phonemes from estimation errors

Speech recognition systems generally use delta and delta-delta (velocity and acceleration) coefficients to characterise the dynamics apparent in frame-based representations of speech. These coefficients can be thought of as the errors of simple predictors. This paper describes the use of error coefficients derived from more advanced (and accurate) forms of prediction and interpolation. Both overall recognition accuracy and the detailed confusions observed are compared with those of the ‘traditional’ methods. The task used is speaker-independent phoneme recognition using a subset of the TIMIT database, and four different speech representations. The error coefficient performance on this task appears to be directly related to the robustness of the estimator used, with the best of the new methods out-performing delta-delta coefficients by around 10%.