Speaker identification based on frequency distribution of vector‐quantized spectra

At present, one of the most important problems in speech recognition and speaker recognition is the extraction of individual information from the speech waveform. This paper describes the extraction of individual information by the vector‐quantization and the text‐independent speaker information based on that method. A feature vector is proposed for the first time which is the quantized distribution by the frequency of the vector‐quantization code to represent the individual features of the speaker. The properties of the feature vector are investigated, and effectiveness is verified by an actual speaker‐identification experiment. The quantization distribution is a feature representing the distribution density in the space for the acoustic features, e.g., the spectrum uttered by the individual. As the acoustic feature parameters, the cepstrum for stationary part, and the change of the cepstrum, are used to construct the quantization distribution. The identification rates are compared. As a result of the identification experiment for 10 speakers, an identification rate of 100 percent was achieved by the quantization distribution of cepstrum for 10 input words, which are different from the training samples. In the experiment using 200 speakers, an identification rate of 88 percent was achieved for the first candidates, and a cumulative identification rate of 95 percent was achieved for up to the second candidate.