CREPE / torchcrepe

About

Deep CNN for monophonic pitch (F0) estimation from raw audio. Original in Keras/TensorFlow by Kim et al. (NYU MARL, 2018). PyTorch port (torchcrepe) by Max Morrison is better maintained and recommended. 360 pitch bins from C1 to B7 at 20-cent resolution, 10ms frame rate. Much more accurate than classical DSP pitch trackers (pYIN, SWIPE). Essential building block for monophonic instrument transcription.

Relevance

Perfect for banjo solo pitch tracking. Banjo is primarily monophonic (melody + drone), and CREPE's 20-cent resolution captures slides and bends. The torchcrepe package is pip-installable and actively maintained. Combine with onset detection for note-level segmentation, then string/fret assignment for tab output.

Mentions

• ../sources/2018-02-16-crepe — paper
• ../entities/basic-pitch — polyphonic alternative
• ../concepts/self-supervised-audio-representation — newer self-supervised pitch trackers (PESTO, 2025)

Links

• Original: https://github.com/marl/crepe (1.4k stars)
• PyTorch: https://github.com/maxrmorrison/torchcrepe (516 stars)
• Install: pip install torchcrepe