| 摘要 |
A method and apparatus for computer speech recognition based on time-domain phoneme waveform identification. In Mainwave-Ripple Model (MRM) analysis, the waveform is located in a short interval called a frame, and waveform structural features are located and measured, to form a waveform analysis array in terms of fine-structure parameters and main-structure parameters. Analysis can also derive multi-frame pattern analysis arrays. In training mode, phoneme reference arrays are formed by combining a known most-probable analysis array with its corresponding phoneme symbol. In recognition mode, unknown input signal analysis arrays are compared with prestored reference arrays, whereby a best-match recognition is made. Selectable levels of processing provide selectable speed versus accuracy, in terms of protowave, phonode, or phoneme recognition. A computer program storage device readable by a computer system for implementing the method is included.
|
