CREPE

CREPE (Kim et al., arXiv 1802.06182; ICASSP 2018) is the model that convinced the MIR community neural pitch tracking was worth it. It takes a 1024-sample slice of raw time-domain audio at 16 kHz, runs it through six convolutional layers to a 2048-dimensional representation, and outputs a 360-bin distribution spanning six octaves (C1 to B7, 20-cent bins, 32.70 to 1975.5 Hz). The pitch estimate is a weighted average over the bins, plus a confidence measure for voicing. The authors trained on a large synthetic music dataset and reported state-of-the-art monophonic pitch tracking, outperforming pYIN and other classical baselines (author-reported). The marl/crepe repo (MIT) ships a pre-trained model with pip-installable inference, which is why CREPE became the default baseline that later trackers like RMVPE compare against. One honest caveat from later literature: it was trained on synthetic music with no real voice samples, so voice-specific retraining is sometimes advised.

• Reported state-of-the-art monophonic pitch tracking at publication, beating pYIN, SWIPE and RAPT (author-reported)
• Widely reused as the evaluation baseline for later neural pitch estimators
• Trained on synthetic music only; one cited analysis notes this can limit performance on real voice signals without retraining
• No independent replication of the headline figures found