Pitch estimation is a basic and significant topic of research in acoustic signal processing. However, because of the interference of noise and instrument, especially the strings with a harmonic structure, pitch estimation is difficult to be researched. And the comb structure formed by the fundamental frequency and its harmonic partials in the spectrum is the important distinction between the pitch and the white noise or other colored noises.
Considering the half-and multi-frequency errors prone to occurring in pitch estimation, this research proposes a pitch estimation method based on harmonic salience in the hope of making the best of the pitch harmonic structure to control these problems. On observing the spectrum with fundamental frequency presenting, it can be found that the harmonic partials are always the local peaks within one fundamental frequency width in most instances. And within the double fundamental, there may be two or more competitive peaks. Although there is only one peak within the half fundamental, the advantage of the harmonic is not significant compared with the magnitudes of the surrounding frequency. That is, if the hypothetical pitch does not correspond with the truth, the significance of the partial will decrease considerably within a hypothetical pitch width.
This method utilizes the characteristics mentioned above to enhance the noise immunity and suppress the half-frequency errors and multi-frequency errors effectively. It uses an inhibiting factor defined as eq. (1) to suppress the values of the spectrum peaks when calculating the harmonic salience summation (eq. (2)) of half- and multi-fundamental frequency.
(1)
(2)
Compared with the method using the original spectrum, this approach can reduce 80% of the half-and multi-frequency errors. As shown in Fig. 1, experiments on various noised data of the TIMIT database demonstrate the robustness compared with other state-of-the art pitch estimation approaches.
Fig. 1 Results of the algorithms on noisy speech data (Image by SONG).
The research has been published on Electronics Letters (Vol. 49, No. 23, pp.1491-1492,07 November 2013).
Contact:
SONG Liming
Institute of Acoustics, Chinese Academy of Sciences, Beijing, 100190, China
Email: songliming@hccl.ioa.ac.cn