Speaker verification based on comparing normalized spectrograms

Voice is one of the primary biometrics and can be used to identify a person. By comparing the similarity of two spectrograms, one is able to investigate if the two voice samples were originated from a same speaker. In this paper a method to normalize the spectrogram of a given voiced sound is presented. Through this normalization process, the differences between the spectrograms of two voice samples due to factors such as speed of utterance, loudness, and the frequency responses of the recording devices are removed. As a result, it is expected that normalized spectrograms reflects mainly the tonal characteristics of its speaker and is more suitable for voice comparison than the original spectrograms. For speaker verification, if the correlation coefficient between two normalized spectrograms is higher than a threshold value, the two sound samples will be regarded as originated from a same speaker. In the experiment, voice samples from 36 males and 33 females were collected. The proposed method was used to conduct speaker verification. When only one sentence (around 6 to 10 Chinese characters) was used, 94% or higher speaker verification accuracy was achieved. The accuracy rate increased when more sentences were included. When 7 sentence were used for comparison, the accuracy rates exceeded 99%.