PitchTracker

On-device CREPE for Swift — the pitch of a recording every 10 ms, in cents, hertz and MIDI note number, with the network's confidence that the frame is pitched at all. CREPE (Kim, Salamon, Li and Bello, ISMIR 2018; MIT) runs on Core AI; the framing, both decoders and the confidence are the original crepe pipeline in Swift. Nothing leaves the machine.

import PitchTracker

let tracker = try await CrepeTracker(contentsOf: assetURL)            // tune-crepe-full-float32.aimodel
let track = try await tracker.track(contentsOf: url)                  // any file AVFoundation decodes
for frame in track.voiced() {                                         // confidence ≥ 0.5
    frame.time, frame.frequency, frame.cents, frame.midi, frame.confidence
}
let viterbi = try await tracker.track(contentsOf: url, method: .viterbi)   // crepe's viterbi=True

What it is exactly

The network is CREPE's, from torchcrepe's port of the original Keras weights (full, 22 M parameters, and tiny, 0.5 M): 1024 normalised samples at 16 kHz in, 360 sigmoid activations out over bins 20 cents apart from 1997.38 cents above 10 Hz (31.70 Hz) to 9177.38 (1975.5 Hz). It is exported at a fixed batch of 256 frames; the host pads the last batch.

Everything else is marl/crepe's core.py, not torchcrepe's, because torchcrepe dithers its cents with random noise and squashes the activations through a second sigmoid, and the original is deterministic: centre zero-padding of 512 samples, a frame every 160 samples (10 ms), each frame normalised by its mean and biased standard deviation clipped at 1e-8 (CrepeFrontEnd); cents as the activation-weighted mean of the nine bins around the argmax (PitchDecoder.localAverage, crepe's default); or viterbi, a 360-state HMM on the argmax bins with a triangular transition of width 12 and self-emission 0.1, then the local average around each path bin; confidence as the largest activation; frequency 10·2^(cents/1200).

The Viterbi decoder reproduces hmmlearn's C++ exactly, tie rule included — the last index on a tie during backtracking, the first for the final state — because unvoiced stretches are full of equal-probability bridges and the tie rule decides the path.

A 16 kHz file is read sample for sample; other rates go through AVAudioConverter at its best quality (upstream uses resampy), which is a resampler difference, not a model one.

Measured against upstream

Tools/export_crepe.py writes both assets and, per clip, the 16 kHz audio, the activations, both decodings and the confidence from the vendored original pipeline; swift test compares. The original Keras crepe package was run on the same clips as a second reference.

Stage	Agreement
Keras crepe vs the exported network (3 clips)	145–151 dB PSNR on the activations, argmax identical
Vendored decoders vs crepe.predict	0.001 cents (float32 vs float64)
Framing, 3 clips	> 120 dB vs crepe's get_activation frames
Network on Core AI, full, 4 clips	141–145 dB on the activations
Whole clips end to end, 7 clips × 2 decoders	max 0.001 cents on every voiced frame; confidence to 4e-6
Network on Core AI, tiny, 2 clips	145 dB
Viterbi path vs hmmlearn	identical on every clip and on random sequences

Synthetic tones as ground truth: a steady A3 + 13 cents reads A3 + 11 cents at confidence 0.94; C4 with ±40-cent vibrato at 6 Hz reads a mean of 60.017 with a 73-cent swing; a two-second glide from E2 to E3 reads 41.17 → 50.85 (expected 41.2 → 50.8). On an M3 Max the full model takes about 0.2 s per ten seconds of audio.

The models

Not bundled. Export them from torchcrepe's weights, or download:

uv run Tools/export_crepe.py --capacity full --install     # and --capacity tiny
hf download arraypress/tune-crepe --local-dir models

The command-line tool is tune.

Requirements

• macOS 27+ (Core AI), Apple silicon
• Swift 6

License

MIT — see LICENSE. CREPE and its weights are MIT (Kim et al.); the PyTorch weights are torchcrepe's (Max Morrison, MIT).