CREPE (safetensors)

Convolutional pitch tracker from Kim et al., ICASSP 2018. Original implementation: marl/crepe.

This repository hosts the five published capacities (tiny, small, medium, large, full), converted from the upstream TensorFlow weights to safetensors for use with the candle ML framework via candle-crepe.

Files

File	Capacity multiplier	Approx. size
tiny.safetensors	4	1.9 MB
small.safetensors	8	6.2 MB
medium.safetensors	16	23 MB
large.safetensors	24	49 MB
full.safetensors	32	85 MB

Tensor layout follows PyTorch conventions. Convolutions are stored as Conv1d (out, in, kernel), the dense classifier as Linear (out, in), and BatchNorm parameters are split into weight, bias, running_mean, running_var.

Names:

conv{i}.conv.{weight,bias}                          i in 1..=6
conv{i}.bn.{weight,bias,running_mean,running_var}   i in 1..=6
classifier.{weight,bias}

Provenance

Converted from the bundled .h5 weights of the crepe PyPI package using scripts/export_safetensors.py.

Parity

Each capacity reproduces the reference TensorFlow forward pass to within 1e-4 max absolute difference on the per-bin activation matrix and on decoded pitch. Verification runs in scripts/pytorch_parity.py and in the Rust integration tests under candle-crepe/tests/.

Citation

@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}
}

License

Same as upstream CREPE: MIT, Copyright (c) 2018 Jong Wook Kim. See LICENSE.