Algorithm and DSP-implementation for a speaker-independentsingle-word speech recognizer with additional speaker-dependent say-infacility

摘要

We present a novel system implementation concept forsmall-vocabulary single-word speech recognition. The system canrecognize speaker-independently a number of preprogrammed words, andadditionally it can perform recognition of a number of words which havebeen trained by simple say-in. In contrast to other implementations, oneand the same algorithm is used for both speaker-dependent andspeaker-independent recognition. This implies that in our proposedsystem the fixed (speaker-independent) part and the personalized(speaker-dependent) part of the vocabulary can be activatedsimultaneously, which is mandatory for the use in mobile communicationsystems. Compared to state-of-the-art speaker-dependent recognitionsystems, the RAM capacity for storing the personalized part of thevocabulary is drastically reduced. The recognition rate is comparable tothat of a fully speaker-independent system. A DSP implementation of thespeech recognition system is described