摘要 |
Wavelet transform means 35 decomposes digitized input speech templates into energy vectors represented by coefficients assigned to fundamental building blocks; a plurality of the wavelet coefficients represents speech signals in time, scale and frequency domains. A profile of the input speech utterance is then constructed by accumulating the energy vectors in blocks. At the same time, the transient response of the energy vectors are also obtained 50. The transient response is the difference in the magnitude of an energy vector and that of the adjoining block. With the transient response, the input speech templates are aligned with those of the reference templates in the library 65 without having to time warp the time axis of the respective templates. The distance between the transient response of a test template and that of a reference template is then checked 60. If a threshold is exceeded, then there is no matching. <IMAGE>
|