FCPE — ONNX

FCPE (Fast Context-based Pitch Estimation, CNChTu) from torchfcpe 0.0.4, bundled model fcpe_c_v001: a conformer on a 128-band log-mel.

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
`fcpe.onnx`	FCPE	10.8 M	47.8 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/fcpe-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, "fcpe.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

FCPE

Input: audio float32 [1, N] — 16 kHz mono
Output: salience float32 [1, T, 360], T = N // 160 + 1 frames, frame i centred at i × 10 ms + 5 ms (torchfcpe's uncentred STFT: a 432-sample pad before a 1024-sample window).
Front end: torchfcpe's mel (432-sample reflect pad, uncentred STFT n_fft 1024 hop 160, 128 Slaney mels 0–8000 Hz, log clamp 1e-5, last frame repeated) is inside the graph.
Exactness vs the original PyTorch pipeline: 4 clips (speech of three lengths + 3 s of noise, 2696 frames): max |ONNX − PyTorch| salience 3.84e-06, max pitch difference 0.00029 cents (1 cent = 1/100 semitone), voicing agreement 100.0 %.
Voicing threshold: 0.33 (chosen on PTDB-TUG's dev speakers, lowest F0 frame error; the upstream default is 0.006).

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

condition	RPA ↑	RCA ↑	GPE ↓	VDE ↓	FFE ↓	VR ↑	VFA ↓
clean	91.83	92.57	1.14	2.51	2.77	94.49	1.58
white noise 10 dB	90.95	91.79	1.26	2.67	2.95	93.74	1.55
white noise 0 dB	89.10	90.02	1.43	3.07	3.39	92.46	1.67
pink noise 10 dB	90.04	91.12	1.56	2.95	3.30	93.48	1.83
pink noise 0 dB	84.69	85.99	1.97	4.36	4.78	89.63	2.47

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	0.0219	358 MB
4	10 s	0.0068	357 MB
1	30 s	0.0214	427 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/fcpe-onnx