摘要 |
PROBLEM TO BE SOLVED: To reduce the calculation required for vector quantization as much as possible and obtain a correct quantized result when a feature vector obtained by voice analysis is vector-quantized by using codebooks. SOLUTION: The codebooks A, B... corresponding to subwords (e.g. phonemes) are created. A universal codebook U having some representative codes CU1, CU2, CU3, CU4 is prepared. The vector quantization calculation for an inputted feature vector V is implemented by using the universal codebook U and the representative codes CU1, CU2, CU3, CU4. When the code (e.g. CU2) is determined as a result of vector-quantization, the vector-quantization calculation is implemented by using only codes existing within a range A1 for implementing the vector-quantization of the codebooks A, B... corresponding to the subwords preset to the code CU2. |