UAI-Software/crepe-onnx

CREPE Pitch Detection (ONNX)

ONNX export of CREPE (Convolutional REpresentation for Pitch Estimation), a monophonic pitch (F0) tracker. Two capacity tiers:

File	Size	Notes
crepe_onnx_tiny.onnx	~1.96 MB	Fast, slightly less precise
crepe_onnx_full.onnx	~89 MB	Reference-quality

Re-hosted by UAI-Software for Kitlig's Vocal Studio pitch-tuning piano roll. Verified end-to-end before use: synthetic sine waves at 110, 220, 440 and 880 Hz all decoded to within ~1% of ground truth at 0.80–0.97 confidence.

Graph contract

• Input: frames, float32, shape [n_frames, 1024] — 1024-sample (64ms @ 16kHz) audio frames, one row per frame. Framing/hop-windowing is the CALLER's responsibility; the graph is a single forward pass over already-framed audio.
• Output: probabilities, float32, shape [n_frames, 360] — independent sigmoid activations (NOT softmax) over 360 pitch-class bins spanning 20 cents each.
• Bin → Hz: cents = 1997.3794084376191 + bin * (7180/359); hz = 10 * 2^(cents/1200).
• Decode with a weighted local average around the argmax bin (± a few bins), not argmax alone, for sub-bin cents precision — see the original CREPE paper's to_local_average_cents.
• Standard hop is 10ms (160 samples @ 16kHz), CENTERED frames (pad ~512 samples on each end so frame 0 is centered at t=0).

Provenance and license

Full chain is MIT:

1. CREPE — Jong Wook Kim et al., https://github.com/marl/crepe — original TensorFlow weights.
2. torchcrepe — Max Morrison, https://github.com/maxrmorrison/torchcrepe — PyTorch port; weights converted from the original TensorFlow weights via MMdnn (same weights, ported format).
3. ailia-models CREPE export — axinc-ai, https://github.com/axinc-ai/ailia-models/tree/master/audio_processing/crepe — PyTorch → ONNX (opset 11) export of torchcrepe's tiny/full variants. That directory's own LICENSE carries both the original crepe (Kim, 2018) and torchcrepe (Morrison, 2020) MIT texts, reproduced verbatim in this repo's LICENSE file.

These exact .onnx files are unmodified from the ailia-models export (verified by file size and graph structure). See LICENSE for the full MIT texts and required attribution.