The classification of noisy speech signal into voiced, unvoiced, and silence provides a preliminary acoustic segmentation for audio signal processing applications, such as digital coding, speech enhancement, and identification. The proposed technique employs a two-staged structure and uses the normalized segmental energy, normalized partial sum of autocorrelation coefficients, and spectral envelope pattern matching as the measurement criteria to classify the voiced and unvoiced speech segments from background noise. The reference noise pattern is updated by incoming silent segments and the classifier can work in real time. Simulation results have shown that this technique is robust even if the audio signal is corrupted by heavy broadband noise.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.