FredrikKarlssonSpeech/torchcrepe-onnx

torchcrepe (ONNX)

ONNX export of maxrmorrison/torchcrepe, a PyTorch implementation of CREPE (Kim et al., 2018), a convolutional neural network for monophonic fundamental frequency (F0) / pitch estimation directly from the raw waveform.

This is an unmodified re-export of the upstream tiny/full checkpoints — same weights, same architecture, only the serialization format changes. It is the same conversion already used in production by superassp, an R package for acoustic phonetic analysis.

Files

File	Precision	Size	Notes
onnx/tiny.onnx	fp32	1.96 MB	recommended
onnx/tiny_fp16.onnx	fp16	1.00 MB	95.6% bin agreement vs fp32
onnx/tiny_int8.onnx	int8 (dynamic)	0.52 MB	82.7% bin agreement vs fp32
onnx/full.onnx	fp32	89.0 MB	recommended
onnx/full_int8.onnx	int8 (dynamic)	22.3 MB	78.8% bin agreement vs fp32, experimental

full_fp16.onnx is intentionally not included — naive fp16 casting broke the full model (only 37.8% pitch-bin agreement with fp32 on the validation audio; see CONVERSION_REPORT.md). Use full.onnx (fp32) or, if size matters more than accuracy, full_int8.onnx.

I/O spec

input:  "frames"        float32 [num_frames, 1024]   # one 1024-sample @16kHz window per row
output: "probabilities"  float32 [num_frames, 360]     # sigmoid activation per pitch bin, pre-decode

Framing and decoding are not part of the graph — you must do these yourself, exactly as the original torchcrepe package does in Python:

1. Resample audio to 16 kHz mono.
2. Slice into overlapping 1024-sample windows (default hop 160 samples = 10ms).
3. Run the model to get a [num_frames, 360] activation matrix.
4. Decode each row to a pitch: convert the argmax (or weighted-argmax, or Viterbi-smoothed) bin index to cents (cents = 20 * bin + 1997.3794084376191) then to Hz (hz = 10 * 2^(cents/1200)).
5. Use the activation at the chosen bin as a periodicity/confidence score for voicing decisions.

See maxrmorrison/torchcrepe's decode.py and convert.py for the exact reference math.

Quick start (Python, onnxruntime)

import numpy as np
import onnxruntime as ort

sess = ort.InferenceSession("onnx/full.onnx", providers=["CPUExecutionProvider"])
frames = np.zeros((1, 1024), dtype=np.float32)  # replace with real 16kHz windows
(probabilities,) = sess.run(None, {"frames": frames})
bins = probabilities.argmax(axis=1)
cents = 20 * bins + 1997.3794084376191
hz = 10 * 2 ** (cents / 1200)

Validation

Re-exported directly from the upstream .pth checkpoints, opset 17. Verified numerically identical (max abs diff ~1e-6, float32 noise floor) to the original PyTorch model, and byte-for-byte identical inference output to the crepe.onnx files already deployed in superassp. Full methodology and quantization drift numbers in CONVERSION_REPORT.md.

License

MIT, inherited from upstream maxrmorrison/torchcrepe. Please cite:

@inproceedings{kim2018crepe,
  title={Crepe: A convolutional representation for pitch estimation},
  author={Kim, Jong Wook and Salamon, Justin and Li, Peter and Bello, Juan Pablo},
  booktitle={2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={161--165},
  year={2018},
  organization={IEEE}
}