| 摘要 |
PURPOSE:To improve the voice recognition rate by generating error vectors by finding the difference between each code vector of a standard speaker code book and adaptive learning data, weighting the error vectors taking the positional relationship between the code vectors and the learning data into consideration and vector quantizing the learning data with a small quantization error. CONSTITUTION:In a speaker adaptive method without teacher of the code book for voice recognition employing vector quantization, error vectors are generated by taking the differences between each code vector of a standard speaker code book and all adaptive learning data. Then, the error vectors are weighted by taking the positional relationship between the code vectors of the standard speaker code book and the adaptive learning data into consideration. The weighted error vectors are added to the code vectors and the standard speaker code book is corrected. By repeating the correction by the weighted error vector, the quantization error is reduced. |
