Automatic extraction of acoustic prototypes for large vocabulary speech recognition by using speaker-independent features

A description is presented of AUTOSEGM, a novel method for automatic extraction of acoustic prototypes for large-vocabulary speech recognition systems based on diphone-like subword segments. AUTOSEGM compares favorably with previous methods for automatic extraction of diphone-like prototypes in that it does not require a set of training templates derived from another talker to bootstrap the new talker´s prototypes. AUTOSEGM exploits only general speaker-independent acoustic/phonetic knowledge, in the form of a general language model and standard acoustic parameters (cepstral derivative, smoothed energy contour). The core of AUTOSEGM is a self-segmentation procedure: the speech material is segmented into pseudosyllables, within which some occurrences of the diphone-like segments are located. A recognition performance of about 78% correct for the top candidate was achieved in a large-vocabulary isolated word recognition task