User:Trevor Law

From MIREX Wiki

As personal project I am developing a multipitch detection system that can run on a low-cost DSP. Been doing in free time for ~1 year. Current iteration is a modification of time-domain algorithm linked here, substantially modified for computational efficiency: https://courses.physics.illinois.edu/phys406/NSF_REU_Reports/2005_reu/Real-Time_Time-Domain_Pitch_Tracking_Using_Wavelets.pdf

Do not like how poorly it performs with multiple strong voices and clipping distortion. Tried doing some FFT+harmonic partial correlation but it appears very unstable. I am signing up for account to access your test datasets. I need to evaluate what I am doing on a faster-to-debug platform than the chip itself, so plan to make fixed-point versions of FFT etc. in Octave or Python and then running the data on them. Need to determine if problems are due to fundamental issues with algorithm, or various sacrifices forced by the hardware (limited FFT-length, fixed-point dynamic range reduction, ADC resolution, etc.).

I appreciate your group's time and effort to generate ground-truth annotated dataset.