( Vocal Pitch Tracker )

Goals

My undergraduate research partner and friend Ross Maddox and I were selected to participate in the 2005 REU project at the University of Illinois at Urbana-Champaign, working under Dr. Steve Errede. We chose to develop an improved real-time time-domain pitch tracker for use on monophonic vocal signals. We aimed for high accuracy, low latency, and robustness to noise and missing/weak fundamentals.

The Project

The results of our project were first summarized in this paper submitted to the REU program. Following its submission, I performed a couple additional tests of our algorithm and rewrote our results in the form required for submission to Kalamazoo College as a Senior Individualized Project (SIP). The resulting document is here; N.B. although I am listed as the sole author (as this was required by my college) and wrote this document alone, Ross was equally responsible as I for the development of the algorithm. An abstract of our work is provided below.

Abstract

A pitch tracker based on the fast lifting wavelet transform (FLWT) is developed in Matlab and C++. The algorithm is designed to function on vocal samples with a range of 90–1500 Hz. With under 25 ms latency, RMS error of under 2 cents (i.e. 1/50 of a semitone) on sinusoidal signals through four octaves is achieved. The algorithm is shown to be robust to noise (up to a SNR of 20–25 dB) and missing/weak fundamentals. Analysis is performed in the time domain using the FLWT. The Haar wavelet is used as the basis for the FLWT, which is shown to be mathematically equivalent to splitting a signal into low-pass and highpass components and down-sampling (generating approximations and details, respectively). Peak detection on successive wavelet approximations is used determine the pitch of vocals.

For a full summary of our work, please read this document.