Dependency of recognition rate on number of words for text‐independent speaker recognition using vector quantization

Hidenori Shimizu, Tetsuo Funada
2008 Journal of the Acoustical Society of America  
Acoustics 08 Paris speaker consists of a feature map. Two kinds of feature vector MFCC and FTTSS of each speaker are used for training the map, and they are quantized into a specific vector on the feature map. The feature FTTSS is used to develop a robust speaker recognition system under noisy condition. Using the map, we conduct speaker recognition(identification and verification) based on vector quantization (VQ) distortion. In particular we examine dependency of recognition rates on number
more » ... utterance words for recognition using the administrative division name of Japan as the utterance words. According to speaker identification experiments, increasing the number of words in recognition more than three words, this system can attain a correct rate of 100% for input speech of 40 speakers. We also examine the influence of the difference of uttering period. Moreover, we show the results of speaker recognition by using Hidden Markov Model (HMM) for comparison with VQ method.
doi:10.1121/1.2935784 fatcat:g6sjywcah5dtdift3hvdtujhf4