audio-health-nano

audio-health-nano

47,094 parameters. 188 KB. Six audio faults, one forward pass, multi-label.

clipped · dropout · dc_offset · hum · bandwidth_loss · hiss

Domain measured / deployment domain tested: measured on ESC-50 clips with synthetic defects, plus 20 s of real office ambience; deployment domain: no real defective broadcast or call audio tested. (Fifth line of the card standard, added 2026-09-02: a number is only as good as the domain it was measured in.)

Prior art — this task is established, and this is a small variant of it

Audio defect detection is a solved problem at full model scale. The direct reference point is Higham et al., A No-Reference Model for Detecting Audio Artifacts using Pretrained Audio Neural Networks, WACV workshops 2022 (Amazon) — 1-second segments classified as No Defect / Hum / Hiss / Distortion / Clicks at 0.986 balanced accuracy, using a modified PANN-CNN6 (~4.8M parameters) on log-mel spectrograms. Broadcast QC tools have detected hum, clipping and dropouts for decades, some of it patented.

Their 0.986 is the reference for what a full-size, pretrained model does. This does not match it and does not claim to. What differs:

Use theirs if you want accuracy. Use this if you need something that runs in 188 KB on a CPU thread, and you want to see which heads actually earn a model.

Measured — ESC-50, split by the dataset's own folds

hiss and bandwidth_loss do not reliably beat a single threshold. Both are broad spectral-shape changes that one statistic captures well. Across repeated runs hiss has scored anywhere from +0.015 to −0.068 against a standard-deviation threshold — that is, sometimes worse than the threshold. Treat both as free riders on a shared forward pass, not as reasons to use a model. If you only need those two, use a threshold.

The cross-corpus test, and what it exposes

Trained on ESC-50, evaluated on 20 s of real office ambience from a webcam microphone — a genuinely different source. The scalar baseline here is a threshold fitted on ESC-50 and transferred, which is what a deployed threshold actually is.

hum scoring below chance is a labelling artefact, not a model failure

The office recording already contains mains hum. Measured against its own broadband noise floor:

Every "no hum" example in that test is a recording with 30 dB of hum in it. The head is detecting hum that is genuinely present, against a label saying it is absent. In any environment with mains equipment, the negative class for hum does not exist, and a cross-corpus number measured there is meaningless.

This generalises: an operation the acquisition environment already performs cannot be detected as applied-versus-not. The same effect kills sharpening and denoising detectors on webcam frames, because the camera's own pipeline already sharpens and denoises.

Which heads to use

• dropout — the only head passing a strict cross-corpus gate. Use it.
• dc_offset, clipped — strong in-corpus, positive but modest across corpora. Use with a validation set from your own material.
• hum — use on media files where a clean negative exists. Do not evaluate it on live room audio from a mains-powered environment.
• hiss, bandwidth_loss — a threshold does nearly as well. Use them because they come free with the forward pass, not because they need a network.

Scope

For: media QC triage, archive ingest checks, flagging recordings for human review, and monitoring capture chains on hardware too small for a pretrained audio network.

Not for:

• Not speech quality, intelligibility, or MOS estimation.
• Not a judgement about content or speakers. It sees a 64×64 spectrogram.
• Not a replacement for a full-size detector where accuracy matters more than size — see the prior art above.
• Not calibrated for music mastering or any perceptual quality decision.

Known failure modes

1. Faults are synthesised onto real recordings. Real clipping involves the whole analogue chain; real dropouts have codec-specific concealment.
2. 1 second at 16 kHz. Content above 8 kHz is invisible; faults shorter than a few ms may be missed.
3. dropout recall is 0.719 in-corpus — it misses about a quarter of cases while being the most transferable head.
4. Environment contamination, as above.
5. No pretraining, so no AudioSet-style semantic prior. It reads spectrogram structure only.

Usage

import numpy as np, cv2, onnxruntime as ort
from scipy import signal as sg

SR = 16000
FAULTS = ["clipped", "dropout", "dc_offset", "hum", "bandwidth_loss", "hiss"]

def spec(x):                                  # x: 1 s of mono float audio at 16 kHz
    f, t, S = sg.stft(x, SR, nperseg=256, noverlap=192)
    P = np.log10(np.abs(S) + 1e-8)
    P = np.clip((P + 8) / 8.0, -1, 3)         # ABSOLUTE level preserved -- see note below
    return cv2.resize(P.astype(np.float32), (64, 64), interpolation=cv2.INTER_AREA)

so = ort.SessionOptions()
so.intra_op_num_threads = 1
so.add_session_config_entry("session.intra_op.allow_spinning", "0")
sess = ort.InferenceSession("audio_health.onnx", sess_options=so,
                            providers=["CPUExecutionProvider"])

probs = 1 / (1 + np.exp(-sess.run(None, {"spectrogram": spec(x)[None, None]})[0][0]))
print({f: round(float(p), 3) for f, p in zip(FAULTS, probs)})

Do not per-window standardise the spectrogram. Sibling models in this family do, and it deletes absolute level — which is exactly what clipping is. With standardisation the clipped head sits at 0.503, chance; keeping absolute level takes it to 0.828. The scaling above is deliberate.

Deployment note: cap the ONNX Runtime thread pool

ONNX Runtime sizes its intra-op pool to the core count and those workers spin-wait between inferences, burning about 1.9 cores to serve 0.28 ms inferences. With intra_op_num_threads=1 and allow_spinning=0, measured idle CPU dropped from 192% to 16.5% of one core, throughput unchanged.

Training

7,000 train windows from ESC-50 folds 1–3, 2,400 test from folds 4–5 · faults sampled independently at p=0.30 so mixtures occur naturally · log-STFT (nperseg 256, noverlap 192), absolute level scaled to a fixed range, resized to 64×64 · 4 conv layers (16→32→48→64) · BCE with positive weighting capped at 10× · Adam 3e-3, 20 epochs.

Same architecture as resoajoe/alarm-nano, resoajoe/camera-health-nano and resoajoe/blockgrid-nano, unchanged.

Verification

ONNX vs PyTorch, both CPU, 256 inputs: max relative logit difference 2.8e-07, 100% threshold agreement.

What "scalar baseline" means on this card

Every margin quoted here is against a stated baseline, because a margin without one is not a measurement. The baseline is the best single-threshold classifier over ten cheap statistics, fitted optimistically:

mean · std · lapvar (Laplacian variance) · hf (high-frequency energy ratio) · grad (Sobel magnitude) · entropy · centre_edge · radial_slope · row_fft_peak · col_fft_peak

The last four are spatially aware, and they were added after an earlier six-statistic baseline — all global aggregates — was found to systematically overstate model value on spatially structured tasks. A baseline that cannot see where anything is loses to a CNN by default. On one test task that flaw inflated an apparent margin from +0.060 to +0.261.

Two different questions are asked with it:

• in-sample — the threshold is fitted on the same data it is scored on. Deliberately generous. Answers is there structure here beyond a low-order statistic?
• transferred — the threshold is fitted on the training corpus and applied unchanged to the target. Answers what should I actually ship?, and they disagree: on one task the in-sample figure was 0.954 and the transferred figure 0.565.

Provenance

ESC-50 (public, CC BY-NC) plus 20 s of the author's own office ambience. No speech, no identifiable people, no personal data.