Building speakers' vowel models and its application in text independent speaker verification

In text-independent speaker verification, we compare two sets of sentences with different text content for their tonal similarity to determine if they were due to the same speaker. Since the sentences are different, we may not have matching words to compare. However, the sentences are constructed from the same set of phonemes of the language used, including vowels and consonants. Generally speaking, vowels are fewer in number, but are the more significant parts of a sentence in terms of duration and loudness, very suitable to be used for tonal comparison. In this paper, we first built spectral models for the simple vowels in Mandarin Chinese. Then we applied the models to analyze two given sets of speech sentences, detecting the various simple vowels in the sentences, and used the detected vowels to build a tonal model for each speaker. After that, we proceed to compare the two tonal models to determine the probability that the two speakers are indeed the same person.