This paper employs time-frequency transformation technology and the Transformer self-attention mechanism to construct a timbre detection model, combined with a source-filter model to perform quantitative analysis of the formant parameters of five vowel categories. The experiment uses the Hann window-optimized STFT algorithm to extract time-frequency features and employs a multi-head self-attention mechanism to model acoustic feature correlations. The study found that the vocal range of bel canto singing in songs is slightly larger than that of ethnic singing. In the vowels /a/, /u/, and /o/, the F1 and F2 values of incorrect vocalization in the bel canto singing style are smaller than those of standard vocalization. In the vowels /i/ and /e/, the F1 values of oral cavity group vocalization are larger than those of the head cavity group, while the F1 values of standard vocalization are between the two groups. Conversely, in F2 values, the oral group is smaller than the head cavity group, but the standard vocalization F2 values still fall between the two groups.