Apparatus and method adapt to the domain of the unknown model by constructing a hierarchical acoustic models for speech recognition

摘要

(57) how to configure a speech model for speech SUMMARY recognition includes calculating a measure of acoustic dissimilarity of speech units lower. model is applied recursively to the audio unit lower based on the measurement values ​​clustering techniques is calculated that the acoustic differences are hierarchically configured by it is generated automatically. Other levels of the hierarchy with a level due to the application of each of the clustering technology, advances to the specification of a maximum from the specification of the minimum is generated. Matching technique is also disclosed in an unknown domain by using the calibration data of an amount less structure and dimensions of the acoustic model set.