| 摘要 |
It is possible to assure accuracy even at the word boundary by using the phoneme environment dependent acoustic model and suppress increase of the processing amount even when recognizing a continuous speech of large vocabulary. A phoneme environment dependent acoustic model storage unit (3) contains a phoneme state tree, i.e., a tree structure of state series of the preceding phoneme state, central phoneme state, and subsequent phoneme state while collecting a triphone model having the same preceding phoneme and the central phoneme. Accordingly, in order to spread a phoneme hypothesis by referencing the phoneme state tree, the language model stored in the language model storage unit (5), and the word dictionary (4) by the forward matching unit (2), what is necessary is only to spread one phoneme hypothesis regardless of the head phoneme of the subsequent word. Thus it is possible to easily spread a hypothesis regardless of in-word or word-boundary state. Moreover, it is possible to significantly reduce the matching amount when performing matching with the characteristic parameter series from an acoustic analysis unit (1).
|
