A novel recurrent network based pitch detection technique for quasi-periodic/pitch-varying signals

The accuracy of pitch detection algorithms affects the performance of many speech and audio applications such as speech compression, computer music analysis/synthesis and information retrieval of audio signals. In many applications, it is also desired that the algorithms should be robust to background noise. A recurrent network based method is proposed in this paper. Though the proposed method requires more computation compared to some existing methods, it is more accurate and less sensitive to noise. The other advantage is that it requires a smaller time frame to estimate the pitch compared to other methods. Therefore, it is more suitable for tracking the pitch of a pitch-varying signal or a quasi-periodic signal. Both the synthesized and natural tones are used in the computer simulation.