| 摘要 |
A speech recognizer, for recognizing unknown utterances in isolated-word small-vocabulary speech has improved rejection of out of vocabulary utterances. Both a usual spectral representation including a dynamic component and an equalized representation are used to match unknown utterances to templates for in-vocabulary words. In a preferred embodiment, the representations are mel-based cepstral with dynamic components being signed vector differences between pairs of primary cepstra. The equalized representation being the signed difference of each cepstral coefficient less an average value of the coefficients. Factors are generated from the ordered lists of templates to determine the probability of the top choice being a correct acceptance, with different methods being applied when the usual and equalized representations yield a different match. For additional enhancement, the rejection method may use templates corresponding to non-vocabulary utterances or decoys. If the top choice corresponds to a decoy, the input is rejected.
|
