Code for ICASSP 2020 paper "Robust Fundamental Frequency Estimation in Coloured noise"
%%Needs to previously download the fast F0 NLS estimator from https://github.com/jkjaer/fastF0Nls (i.e., the Maximum Likelihood Estimator under a white Gaussian noise assumption). For voicing detection, in the file fastF0Nls.m, in the Step 1 (compute the priors on the pitch and the model ), it is suggested to modify the prior of the model, by assigning a voicing probability, e.g.,
voicingProb = 0.60; %%i.e., the probability of a not-voiced model (either unvoiced speech or silence) is 0.40
logModelPrior = log([(1-voicingProb), ones(1,obj.L)*voicingProb/obj.L]);
%%The ICASSP 2020 results can be obtained from running the file expIterativeforICASSPresults.m (need to add the directory of fastF0Nls)
%%The .pev and .pes files contain the audio samples of Keele database and the peak correlation values used for F0 estimation (assumed here as a "ground truth")
%%In babble noise (and nonstationary noise, e.g., restaurant noise), a better improvement is achieved with using 256 noise spectral envelopes instead of the 16 for the ICASSP submission, when applying the initial pre-whitening (i.e., pre-processing) step.
%%Even if some form of pre-processing (e.g, pre-whitening or speech enhancement) benefits non-parametric pitch estimators (e.g., RAPT and SWIPE), a better accuracy from the iterative F0 (NLS)-AR is observed, specially at low iSNRs, even if the resulting F0 estimates are not smoothed.
%%%In high iSNR scenarios, it might be better not to initially pre-whiten the signal, but still following the iterative estimation (i.e., 1st estimate the pitch and then estimate the AR parameters after an approximate modelled stochastic residual was obtained)
%%At very low SNRs in very non-stationary noise (e.g., babble or restaurant), pre-whitening followed by NLS pitch estimates and applying re-iterative estimation leads to the best performance in terms of FFE (including gross errors and voicing detections) as compared to a joint estimator (in case one is interested in better accuracy) (the difference was around 5%, I think). Although such result was not reported yet in a paper, I hope I can update such results in this repo.