Improvement of pitch detection using a signal specific analyzing wavelet
Yoshifumi Chisaki, Tsuyoshi Usagawa, Masanao Ebata · The Journal of the Acoustical Society of America · 1998
Many methods of a speech enhancement system in a noisy environment have been studied. A wavelet transform performs better in analyzing speech than a fast Fourier transform at the point of frequency resolution when the purpose is focused on the pitch detection. Analyzing wavelets, such as Gabor, are derived from a mathematical approach, which have characteristics of low, high, and bandpass filters. In this paper, a modified wavelet transform is proposed. A signal which has characteristics of the target signal is applied to an analyzing wavelet. This signal-specific analyzing wavelet includes not only amplitude and phase for each frequency but also the relationship of harmonic components and a periodic fluctuation of pitches. In addition, the wavelet transform is arranged to use a signal-specific analyzing wavelet for tracking pitches. This method is useful not only for the pitch detection but also for the representation of the fluctuation. Experiments are performed by using both a Gabor function and a signal-specific analyzing wavelet changing both vowels and a bandlimited noise. As a result, the ratio of pitch detection was improved by comparing a signal-specific analyzing wavelet with a Gabor wavelet transform. The ratio was recovered over 40% in case of SNR=0 dB.