Tip: The estimator described here has recently been generalized to the case of coloured noise. Specifically, fast algorithms for fundamental frequency estimation in autoregressive noise have been found for the case where both the harmonic and autoregressive orders are unknown. Since the estimator described here is a special case of the generalised estimator, we strongly encourage you to check out the following link where both the paper and code can be found: https://github.com/jkjaer/fastF0ArMl.
Periodic signals are encountered in many real-world applications such as music processing, speech processing, sonar, order analysis, and electrocardiography (ECG). Such signals can be modelled as a weighted sum of sinusoids whose frequencies are integer multiples of a common fundamental frequency which in audio and speech applications is often referred to as the pitch. Therefore, an important and fundamental problem in the above mentioned applications is to estimate this fundamental frequency from an observed (and often noisy) data set.
Multiple estimation methods have been proposed in the scientific literature ranging from simple correlation-based methods (such as PRAAT, RAPT, YIN, and Kaldi) to parametric methods (such as subspace methods, filtering methods, harmonic summation, and NLS). Although the parametric methods in general are much more accurate and robust to noise than the correlation-based methods, they typically suffer from a high computational complexity. Consequently, the correlation-based methods remain very popular despite that they require all sorts of heuristic post-processing to give a satisfactory performance.
The nonlinear least squares (NLS) estimator of the fundamental frequency is a parametric method which has been known since 1992 and been shown to be one of the most accurate and robust fundamental frequency estimators, in particular for short segments of data. Unfortunately, the computational complexity of any published NLS algorithm is very high so most people have instead used harmonic summation which is an approximate NLS method. Recently, however, we published a very fast algorithm for computing the NLS estimate and the number of harmonic components, and this git repository contains C++- and MATLAB-implementations of this algorithm. The details of the algorithm can be found in
In addition to these papers, we have also made a one hour video tutorial on fundamental frequency estimation in general and our fast algorithm in particular. The video is published on YouTube.
To get started with the code, please see the examples and the documentation.
Please note that the code only contains a pitch estimator and NOT a pitch tracker. The difference is that a tracker contains a smoothing step on top of the estimator. The smooting step is there to minimise the risk of, e.g., octave errors (aka pitch halving) by smoothing out the estimates produced by the estimator which typically analyse the data on a segment-by-segment basis. Of course, our estimator can also be used inside a pitch tracker. For the best performance, we recommend that the smoothing step by Tabrikian et al. is used. See more in
Alternatively, and much simpler, median smoothing or dynamic programming can also be used. See more in
For white Gaussian noise (WGN), the NLS estimator is the maximum likelihood estimator and is, therefore, asymptotically optimal (in a statistical sense). That is, no other unbiased estimator of the fundamental frequency has a lower variance than the NLS estimator if enough data are observed and the noise is white and Gaussian. In our experience, the NLS estimator is also one of the best estimators for short data segments, and it works well even for situations where a data segment contains as little as only one cycle of the periodic signal. For voiced speech, where the lowest fundamental frequency is typically bigger than 80 Hz, this means that the estimator typically works well down to a segment length of 12.5 ms. Increasing the segment length will, of course, increase the estimation accuracy and the robustness to noise, provided that the signal is approximately stationary.
In our experience, the estimator does typically not break down if the noise is not white and Gaussian. However, if the noise is coloured and has most of the energy at the lower frequencies, then the estimator can suffer from problems with octave errors. In this case, we recommend that the generalisation of the estimator to the case of coloured noise is used. See more here: https://github.com/jkjaer/fastF0ArMl.
The code is published under a BSD 3-clause "New" or "Revised" License. This is a very permissive license that enables the code to be used free of charge in both academia and industry. Besides honouring the license, we therefore also expect that you
Constructive feedback in any form is also very much appreciated. This applies to everything from bugs to data where the estimator fails.
The code has been written by