A Multiple Functions Multiplication Approach for Pitch Extraction of Noisy Speech
Md. Saifur Rahman, Yosuke Sugiura, Tetsuya Shimamura · 2019
This paper addresses a new concept to produce a noise robust pitch extraction function, which is called multiple functions multiplication. A modified version of the autocorrelation function (ACF) (weighted ACF) is recognized as a multiplication of two functions; ACF and inverse of average magnitude difference function (AMDF). Extending the weighted ACF, a three functions multiplication version is derived where the cumulant based ACF (Cum-ACF) is utilized. A cepstrum (CEP) version of the Cum-ACF (Cum-CEP) is also considered instead of the Cum-ACF. The resulting function consists of a multiplication of three functions; ACF, inverse of AMDF, and Cum-ACF (or Cum-CEP). Through experiments, the performance of the proposed method is investigated. It is shown that the proposed method provides an excellent pitch extraction in several noise environments.