Pitches Detection of Mixed Speech using Synchrosqueezing Wavelet Transform
Dandan Chen, Ting Xiao, Weiping Hu, Qing Feng Wu · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
According to the quasi periodicity of speech signal, an approach for mixed speech pitches detection is proposed by using synchrosqueezing wavelet transform. Based on the wavelet transform, the mixed speech signal is transformed into a time-frequency map with high frequency resolution and good energy aggregation. It is found that the time-frequency map can clearly distinguish the pitch ridges of different speakers. Using the continuity between speech frames and the constraints of the maximum energy criterion of time-frequency units, the pitches of each speaker in mixed speech can be well detected by extracting time-frequency ridge; The time-frequency map, which generated by traditional wavelet transform, will produce serious edge effect in the low-frequency part. The time-frequency map generated by traditional short-time Fourier transform has low frequency resolution and poor energy aggregation, which is not conducive to extracting time-frequency ridge. The comparison results show that the approach has good detection ability for the pitches of mixed speech.