Speaker recognition apparatus and speaker recognition method to eliminate a trade-off relationship between phonological resolving performance and speaker resolving performance

A speaker recognition method for a speaker recognition apparatus, the method comprising:

• obtaining a weighted vector distance, which is based on a first weight, between voice feature vectors of a first voice feature vector sequence produced from a voice of a speaker to be registered, clustering the first voice feature vector sequence based on the obtained weighted vector distance, and producing a codebook including a plurality of representative vectors, by a speaker model production part of the apparatus;
• storing the produced codebook in a storage section of the apparatus;
• obtaining a weighted quantization distance, which is based on a second weight, between corresponding ones of the plurality of representative vectors stored in the codebook and each voice feature vector of a second voice feature vector sequence produced from a voice of a speaker to be recognized, quantizing each said voice feature vector of the second voice feature vector sequence based on the obtained weighted quantization distance, and obtaining a weighted quantization distortion, which is based on a third weight which differs from the first weight and the second weight, between the corresponding ones of the plurality of representative vectors stored in the codebook and each said voice feature vector of the second voice feature vector sequence, by a speaker distance calculation part of the apparatus;
• and
• executing speaker recognition based on the obtained quantization distortion, by a recognition part of the apparatus.

A computer readable storage medium having a program stored thereon that is executable by a computer to cause the computer to perform a speaker recognition process comprising:

• obtaining a weighted vector distance, which is based on a first weight, between voice feature vectors of a first voice feature vector sequence produced from a voice of a speaker to be registered, clustering the first voice feature vector sequence based on the obtained weighted vector distance, and producing a codebook including a plurality of representative vectors;
• storing the produced codebook;
• obtaining a weighted quantization distance, which is based on a second weight, between corresponding ones of the plurality of representative vectors stored in the codebook and each voice feature vector of a second voice feature vector sequence produced from a voice of a speaker to be recognized, quantizing each said voice feature vector of the second voice feature vector sequence based on the obtained weighted quantization distance, and obtaining a weighted quantization distortion, which is based on a third weight which differs from the first weight and the second weight, between the corresponding ones of the plurality of representative vectors stored in the codebook and each said voice feature vector of the second voice feature vector sequence;
• and
• executing speaker recognition based on the obtained quantization distortion.