Pitch determination by fundamental-harmonic extraction using MUSIC algorithm

Abstract

Short-term analysis pitch determination was executed by using multiple signal classification algorithm, which is an eigen-based subspace decomposition method proposed by Schmidt (Ph.D. thesis, Stanford University, 1981). The MUSIC spectrum, based on subspace principles, is sharply peaked at the frequencies of the sinusoidal components of speech signals without almost receiving the influence of the noise. Since the harmonic structure of the power spectrum of speech signals becomes unclear in a high-frequency domain, fundamental-harmonic extraction is performed using the band-limited MUSIC spectrum. The influence of the noise decreases further by limiting the frequency domain to be analyzed, and calculation time has been shortened greatly. The IDFT of the logarithmic MUSIC power spectrum exhibits a strong peak at the position equal to the pitch period like the cepstral method. Results of pitch determination for male and female Japanese vowels illustrate that the proposed method is more excellent than the cepstral method and can estimate the pitch frequency not being influenced by the noise.