Speech recognition system and method

Abstract

The invention discloses a voice recognition system and method. In a speech recognition system, a combination of a log-linear model and multiple speech features is provided to recognize unknown speech utterances. The speech recognition system uses a log-linear model to model the posterior probability of language units related to speech recognition. The posterior model captures the probabilities of language units given the observed phonetic features and the parameters of the posterior model. The a posteriori model can be determined with probabilities of word sequence hypotheses given a plurality of phonetic features. Use log-linear models with features derived from small or incomplete data. The speech features used may include asynchronous, overlapping and statistically independent features. Not all features used in training need to be present in detection/recognition.

Description

technical field [0001] The present invention relates generally to speech recognition systems, and more particularly to speech recognition systems that utilize a variety of speech features using log-linear models. Background technique [0002] Speech recognition systems are used to recognize word sequences from unknown speech utterances. In an exemplary speech recognition system, a feature extractor is used to extract speech features, such as cepstra and delta cepstra features, from an unknown utterance to characterize the unknown utterance. A search is then performed to compare features extracted from unknown utterances with models of phonetic units (such as phrases, words, syllables, phonemes, subphonemes, etc.) to compute probability scores for different word sequence hypotheses. In general, the search space is limited by removing unlikely hypotheses. The word sequence associated with the highest score or highest likelihood or highest probability is identified as the unk...

AI Technical Summary

Problems solved by technology

In general, limit the search spac

Images