| 摘要 |
<p>The parallel application of a plurality of vocal tract filters (12) is utilized in a speech recognition apparatus. Each of the filters in this bank of filters has a complex Fourier transfer function that is the reciprocal of a particular vocaltract transfer function corresponding to a particular speech sound. The speech elements identified are phoneme segments of short duration (typically 10 milliseconds), and correspond to phonemes only in the case of time invariant (sustainable) phonemes. In a preferred embodiment, it is assumed that the bank of inverse filters, when designed to correspond with the sustainable phonemes, is capable also of closely matching (in a piecewise fashion) the time-varying (transitional) phonemes. The input of each filter in the bank of filters is connected to the speech waveform input (30). The output of each filter is examined (18) to determine which output has the smallest absolute value. Generally, the filters are designed so that when a filter channel has the smallest output at a given time, there is usually present at the input of the filter bank at that time the waveform of the sound such filter was designed to detect. In a preferred embodiment, the apparatus determines (22) which filter"s output is minimum in absolute value for the greatest total time over a given short interval of time, and the associated sound is specified as the sound present at the input. </p> |
