Pitch estimation using harmonic product spectrum derived from DCT

Estimation of pitch from a given segment of speech plays an eminent role in various speech processing applications, such as speech coding, speech recognition, speaker recognition tasks, speech synthesis, etc. Even though, there are several efficient algorithms, estimation of pitch frequency from speech signals that are severely degraded by noise is still a challenging task. In this paper, we propose a robust framework for pitch estimation using harmonic product spectrum (HPS) derived from discrete cosine transform (DCT) of the signal. This novel method exploits the better decorrelating nature of the DCT spectrum that enables the pitch harmonics to appear sharper in its spectrum. Potentially, this facilitates accurate pitch estimation at lower order of the harmonic product spectrum when compared with DFT-based HPS. Systematic evaluation is carried out to analyze the performance of the proposed method in comparison with some of the successful algorithms, like DFT-based HPS, SIFT, and Cepstrum-based technique. The results clearly show that the proposed algorithm outperforms the other algorithms for speech signals that are severely corrupted by noise (low SNR). The effectiveness of this method for different durations of analysis window, various orders of HPS, and the refinements are also discussed.