A cepstral-domain algorithm for pitch estimation from noise-corrupted speech

Celia Shahnaz, Wei-Ping Zhu, M. Omair Ahmad

Abstract


A study was conducted to develop an accurate algorithm for pitch estimation from noisy speech observations with an aim to significantly reduce the pitch-errors for a wide range of speakers. The study proposed to employ a Discrete Cosine Transform (DCT) based power spectral subtraction scheme for enhancing noisy speech prior to pitch estimation. The de-noised speech was inverse filtered, to yield an output, referred to as the Linear Prediction (LP) residual, to remove the adverse effect of formants. The objective of the proposed method is the introduction of a DCT power cepstrum (DPC) of the LP residual that exhibits a more prominent peak at the true pitch, relative to that demonstrated by the conventional cepstrum of noisy speech.

Keywords


Acoustic variables measurement; Boolean functions; Clutches; Continuous speech recognition; Cosine transforms; Discrete cosine transforms; Electric fault location; Accurate; Adverse effects; Cepstral; Cepstrum; Linear predictions; Noisy speeches; Pitch estimations; Power cepstrum; Power spectral

Full Text:

PDF

Refbacks

  • There are currently no refbacks.