Time delay estimation and adaptive frame length iterations for noise robust pitch extraction

We describe a new pitch tracking method based on a time delay estimation technique and we incorporate an adaptive frame length iteration stage. Given two frames of voiced speech and the time delay between them, we run a linear regression on the unwrapped phase of the quotient of the spectrum of both frames. A weighted linear regression allows us to avoid the effect of phases corrupted by spectral leakage and noise. Iterations adapting the frame length allow us to have better time resolution, less spectral leakage and more noise robustness, avoiding inaccuracies that could come from pitch doubling and jitter.