Frequency Modulation Harmonic and Noise Model for Robust Pitch Contour Extractor

• Main Authors: Chung-Chieh Huang, 黃重傑
• Other Authors: 廖元甫
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/cx69bz

碩士 === 國立臺北科技大學 === 電腦與通訊研究所 === 98 === Robust pitch extraction under high level of noise interference is an important issue for many applications. In this paper, a new frequency modulation harmonic and noise model (FM-HNM) is proposed for robust pitch contour extraction. The main idea is that fundamental frequency is usually not stable and varying over time. Therefore, the set of sinusoid bases in traditional HNM should be modified to reflect this phenomenon. The performance of the method is evaluated using the Keele pitch detection database with artificially added background noise, and compare with several standard methods, including RAPT(Robust Algorithm for Pitch Tracking), voicebox RAPT, YIN, STRAIGHT(Speech Transformation and Representation based on Adaptive Interpolation of weighted spectrogram). It is found from experimental results that when take voiced/unvoiced decision error and traditional GEC(Gross Error Count) into account at the same time, (1) RAPT is most stable in all conditions, (2) STRAIGHT has better performance in clean and white noise situation, but (3) in babble noise case, FM-HNM is better than others. Therefore, the proposed FM-HNM is promising.