Integration of segmenting and nonsegmenting approaches in continuous speech recognition

This paper presents new techniques to improve the recognition rate of time-warping based continuous speech recognition systems. These techniques were implemented in a pre-segmenting recognizer. They reduced the error rate by a factor of three. The proposed techniques address the correction of unexpected segmentation errors, which are a major problem for such recognizers. We implemented a segmentation verifier that uses a recognition-derived spectral similarity measurement to decide if new boundary hypotheses should be proposed. Original and new boundaries are then pursued in parallel until further recognition shows which one to select. Moreover, nonlinear penalties are imposed when a segment´s contraction factor suggests a probable insertion or when poor local similarity suggests a probable substitution. Using only the pre-segmentation, the system recognizes 78% of the test sentences [1,2]. With a nonsegmenting approach, the system performance drops to 73% [3]. With the algorithm proposed here, the system achieves a sentence recognition rate of 92%. These results demonstrate that recognition of continuously spoken sentences can be significantly enhanced by a tighter coupling of the segmentation and recognition stages.