Pitch-Tracking Systems in Auto-Tune

Details

• Built the full real-time auto-tune pipeline in MATLAB, capturing audio at 48 kHz, isolating the vocal band with cascaded Butterworth filters (80 dB stopband attenuation), and decimating in two stages to an 8 kHz analysis rate
• Implemented FFT pitch detection over a Hamming-windowed, 8192-bin spectrum with parabolic peak interpolation, resolving fundamental frequency below the native bin spacing without increasing FFT size
• Implemented time-domain autocorrelation detection with lag bounds restricted to 55-2000 Hz and a correlation-peak prominence threshold, preventing false locks on higher harmonics
• Built the latency and residual-error measurement and reporting tooling behind the comparison, measuring a 2.68x speedup for FFT (0.22 ms per frame versus 0.59 ms) with roughly a third of the variance
• Characterized tracking limits with a continuous chirp sweep, showing autocorrelation collapsing past the 1500 Hz passband edge while FFT tracked into the filter transition band before both failed in the 2 kHz stopband
• Retuned tracked pitch to the nearest 12-tone equal temperament note with exponential smoothing to prevent note flutter, then resampled, Hann-windowed, and normalized each frame for live playback
• Owned the 3-person team's Git repository and integration workflow