CREPE (full, tiny) — ONNX

CREPE (Kim et al., CREPE: A Convolutional Representation for Pitch Estimation, ICASSP 2018) via torchcrepe 0.0.24, both capacities it ships: full and tiny. A CNN on raw 1024-sample frames.

Exported to ONNX with the model's own front end inside the graph: feed 16 kHz mono audio, get the model's own output — no torch, librosa or reimplemented STFT/mel needed, and nothing that can quietly differ from the original. Every export was checked frame by frame against the original PyTorch pipeline (below), and measured on PTDB-TUG speech against the other exports in this family and against pyin.

file	model	parameters	size
`crepe-full.onnx`	CREPE full	22.2 M	89.0 MB
`crepe-tiny.onnx`	CREPE tiny	0.49 M	2.0 MB

Same family: rmvpe-onnx · fcpe-onnx · crepe-onnx · pesto-onnx.

Use

# pip install onnxruntime numpy soundfile soxr huggingface_hub
import soundfile as sf, soxr
from huggingface_hub import hf_hub_download
import importlib.util

repo = "Scicom-intl/crepe-onnx"
spec = importlib.util.spec_from_file_location("pitch_onnx", hf_hub_download(repo, "pitch_onnx.py"))
pitch_onnx = importlib.util.module_from_spec(spec); spec.loader.exec_module(pitch_onnx)

model = pitch_onnx.PitchModel(hf_hub_download(repo, "crepe-full.onnx"))   # threshold= to override voicing
y, sr = sf.read("speech.wav", dtype="float32", always_2d=True)
y = soxr.resample(y.mean(1), sr, 16000)
f0, voiced, confidence = model(y)   # one per 10 ms frame (frame i at i * 10 ms; FCPE + 5 ms); f0 = 0 unvoiced

pitch_onnx.py is numpy + onnxruntime only: it cuts CREPE's frames, decodes the 360-bin salience the way the models ship (salience-weighted mean of the 9 bins around the peak; CREPE/RMVPE bins are 20 cents from 32.70 Hz, FCPE's are evenly spaced 32.7–1975.5 Hz) and applies the voicing threshold.

Details

CREPE full

Input: frames float32 [T, 1024] — raw 16 kHz samples, audio zero-padded 512 each side, hop 160
Output: salience float32 [T, 360], T = N // 160 + 1 frames, frame i centred at i × 10 ms.
Front end: torchcrepe's per-frame mean/std normalisation is inside the graph; cutting the frames is pitch_onnx.crepe_frames (batch long clips by frames).
Exactness vs the original PyTorch pipeline: 4 clips (speech of three lengths + 3 s of noise, 2696 frames): max |ONNX − PyTorch| salience 2.98e-07, max pitch difference 0.00002 cents (1 cent = 1/100 semitone), voicing agreement 100.0 %.
Voicing threshold: 0.61 (chosen on PTDB-TUG's dev speakers, lowest F0 frame error; the upstream default is 0.21).

Accuracy on PTDB-TUG's 16 test speakers (400 utterances per condition), in %:

condition	RPA ↑	RCA ↑	GPE ↓	VDE ↓	FFE ↓	VR ↑	VFA ↓
clean	75.15	76.68	2.68	5.89	6.43	84.63	2.93
white noise 10 dB	74.30	75.90	2.86	5.42	5.98	83.12	1.83
white noise 0 dB	69.51	70.90	2.59	6.44	6.92	76.53	1.12
pink noise 10 dB	73.31	75.03	3.06	6.00	6.59	81.44	2.07
pink noise 0 dB	61.56	62.98	3.24	8.97	9.48	66.77	1.37

CPU speed (onnxruntime, x86 server CPU; RTF = compute seconds per second of audio; peak RSS is the whole Python process, librosa and numpy included):

threads	audio	RTF (compute s / audio s)	peak RSS
1	10 s	2.0914	2473 MB
4	10 s	0.5239	2476 MB
1	30 s	2.0671	2508 MB

CREPE tiny

Input: frames float32 [T, 1024] — as CREPE full
Output: salience float32 [T, 360], T = N // 160 + 1 frames, frame i centred at i × 10 ms.
Front end: As CREPE full.
Exactness vs the original PyTorch pipeline: 4 clips (speech of three lengths + 3 s of noise, 2696 frames): max |ONNX − PyTorch| salience 4.47e-07, max pitch difference 0.00004 cents (1 cent = 1/100 semitone), voicing agreement 100.0 %.
Voicing threshold: 0.58 (chosen on PTDB-TUG's dev speakers, lowest F0 frame error; the upstream default is 0.21).

Accuracy on PTDB-TUG's 16 test speakers (400 utterances per condition), in %:

condition	RPA ↑	RCA ↑	GPE ↓	VDE ↓	FFE ↓	VR ↑	VFA ↓
clean	74.75	76.23	2.40	5.86	6.33	82.73	2.29
white noise 10 dB	73.81	75.33	2.53	5.58	6.07	81.52	1.54
white noise 0 dB	68.66	70.23	2.85	6.70	7.21	75.30	1.07
pink noise 10 dB	72.32	73.82	2.57	5.92	6.41	79.60	1.39
pink noise 0 dB	60.26	61.77	3.42	8.77	9.31	65.92	0.85

CPU speed (onnxruntime, x86 server CPU; RTF = compute seconds per second of audio; peak RSS is the whole Python process, librosa and numpy included):

threads	audio	RTF (compute s / audio s)	peak RSS
1	10 s	0.0906	580 MB
4	10 s	0.0249	576 MB
1	30 s	0.0917	614 MB

Comparison on PTDB-TUG (F0 frame error, %, lower is better)

FFE = (voicing-decision errors + gross pitch errors > 20 %) / all frames — the single number that counts both "is this frame voice" and "how wrong is the pitch". Reference: PTDB-TUG's laryngograph-derived f0 (RAPT, 10 ms). Speakers F01 F02 M01 M02 were used ONLY to pick each method's voicing threshold and frame lag (clean audio, lowest FFE); the table is the other 16 speakers, 400 utterances per condition, scored once. Noise is added at the stated SNR over the whole utterance. Best per row in bold.

condition	RMVPE	FCPE	CREPE full	CREPE tiny	PESTO	pyin (librosa 0.11)
clean	3.09	2.77	6.43	6.33	5.58	7.72
white noise 10 dB	3.18	2.95	5.98	6.07	5.99	8.32
white noise 0 dB	3.56	3.39	6.92	7.21	7.46	16.92
pink noise 10 dB	3.48	3.30	6.59	6.41	6.54	8.36
pink noise 0 dB	4.47	4.78	9.48	9.31	8.99	17.18

Metrics per condition for every method, the chosen thresholds/lags and the CPU timings are in benchmark.json; the harness is export/evaluate.py.

How it was exported

export/export.py (torch 2.8, legacy ONNX exporter, opset 17). The STFT is two conv1ds with fixed DFT kernels (torch.stft's complex output doesn't export), equal to torch.stft to float32 rounding. export/decode.py is the decoder the parity check uses.

Scicom-intl/crepe-onnx