Skip to content

CREPE: A Convolutional Representation for Pitch Estimation

Summary

Deep convolutional neural network for monophonic pitch (F0) estimation operating directly on time-domain audio. Predicts pitch activation across 360 bins covering C1 to B7 (32.7 Hz to 1975.5 Hz) with 20-cent resolution at 10ms frame rate. Significantly more accurate than classical DSP methods (pYIN, SWIPE). Available in multiple capacity sizes (tiny through full). The most widely cited deep learning pitch tracker (669+ citations).

Key Claims

  • Deep CNN on raw audio outperforms hand-crafted DSP pitch trackers by a wide margin
  • 20-cent resolution is sufficient for most music applications (human pitch discrimination is ~5 cents)
  • Model capacity can be traded for speed (tiny model is near real-time on CPU)
  • Training on synthetic data (generated with varying pitch, loudness, timbre) generalizes well to real audio