Direct time domain fundamental frequency estimation of speech in noisy conditions

A new algorithm of direct time domain fundamental frequency estimation (DFE) and voiced/unvoiced (V/UV) classification of speech signal is presented in this paper. The DFE algorithm consists of spectral shaping, detection of significant extremes based on adaptive thresholding, and actual frequency estimation under several truth criteria. We propose a majority criterion for V/UV classification based on the detected frequencies consistency evaluation. Performance of the algorithm is tested on the Speecon database and compared to the Praat modified autocorrelation algorithm. In comparison to the Praat, the results indicate better properties of the DFE for clean speech and speech corrupted by additive noise to SNR about 10 dB. For lower SNR, sensitivity of the DFE to the speech component decreases rapidly while Praat fails to differentiate noise and unvoiced parts of speech from voiced parts.