| --- |
| license: mit |
| tags: |
| - audio |
| - pitch-estimation |
| - crepe |
| - core-ai |
| - apple-silicon |
| - macos |
| library_name: swift-pitch-tracker |
| --- |
| |
| # tune-crepe — CREPE for Core AI (macOS 27) |
|
|
| CREPE (Kim, Salamon, Li and Bello, "CREPE: A Convolutional Representation for Pitch |
| Estimation", ICASSP 2018; MIT) exported to Apple Core AI `.aimodel` assets for on-device pitch |
| tracking. The weights are the original Keras ones via [torchcrepe](https://github.com/maxrmorrison/torchcrepe) |
| (Max Morrison, MIT). Read by [swift-pitch-tracker](https://github.com/arraypress/swift-pitch-tracker) |
| and the [`tune`](https://github.com/arraypress/swift-tune-cli) command-line tool. |
|
|
| ## Files |
|
|
| | Path | What | |
| |---|---| |
| | `tune-crepe-full-float32.aimodel/` | The paper's model, 22 M parameters, 89 MB. `frames [256, 1024]` → `activations [256, 360]`. | |
| | `tune-crepe-tiny-float32.aimodel/` | The small model, 0.5 M parameters, 2 MB. Same shapes. | |
|
|
| Input: 1024-sample frames of 16 kHz audio, each normalised to zero mean and unit (biased) |
| standard deviation, as `crepe.core.get_activation` builds them (centre zero-pad 512, hop 160). |
| Output: sigmoid activations over 360 bins 20 cents apart from 1997.38 cents above 10 Hz. |
| Decoding (local average around the argmax, or Viterbi) is the host's, reproducing `marl/crepe`. |
|
|
| ## Fidelity |
|
|
| The exported network matches the original Keras `crepe` at 145–151 dB PSNR on the activations |
| (argmax identical); on Core AI it runs at 141–145 dB (full) and 145 dB (tiny) against |
| torchcrepe; whole clips decode to within 0.001 cents of `crepe.predict`. |
|
|
| ## Use |
|
|
| ```sh |
| hf download arraypress/tune-crepe --local-dir models |
| tune model install models |
| tune bass_01.wav # E1 −4 cents |
| ``` |
|
|
| Requires macOS 27 (Core AI) on Apple silicon. Converted with coreai-torch 0.4.2 / torch 2.13. |
|
|
