◉ SoundCue_

A visual sound-awareness companion. It listens through your microphone and turns everyday alert sounds — a knock at the door, a baby crying, a siren, an alarm clock — into clear visual signals, entirely on this device. Built as an accessibility aid for Deaf and hard-of-hearing people.

offline · no account · no telemetry trained CNN, in-browser inference awareness aid — not a safety device

01Listen

Live microphone in, visual alerts out. A small convolutional network classifies a 4-second window every 2 seconds, on this device. When confidence is low it says so instead of guessing.

Not listening yet
Alert tiers (starting suggestion — set them for your own home):
0 events

02Try a clip — no microphone needed

Ten real clips from the official fold-5 test split, one per class — held out from all training, though these clips were also inspected during development (see Evidence for the honest read). Play one, classify it, then drag the noise slider and watch confidence fall until the model abstains — that honesty is the point.

03Evidence

Every number below is a measured evaluation of the shipped weights, re-run after the fact — the original training history (logs, checkpoints) was not retained, and the packaged prep_train.py is a reconstructed recipe, not a proven-reproducing run. Data: ESC-50 (10 alert-relevant classes, 400 clips, 40 per class). Official folds 1–3 train (240 clips), fold 4 validation (80), fold 5 test (80). Fold 5 was held out from training, but its clips were inspected during development (interactive noise tests, pipeline fixes, re-evaluation), so read it as a development-exposed benchmark — a measure of this model, not pristine independent validation or household reliability. No threshold or feature was retuned on fold 5 to raise the score. Seed 7. Augmentation on train only: speed 0.9–1.1×, gain, 10–30 dB white noise, time shift, random 4 s crop, mixup, SpecAugment-style masking.

Headline results

Per-class precision / recall (test, 80 clips)

Confusion matrix — test fold (rows: true, columns: predicted)

The browser model was verified against the PyTorch training reference on all 80 test clips: largest probability difference 3.8×10⁻⁷, zero disagreements in predicted class and zero disagreements in the abstain decision at the default threshold. The app computes identical features (log-mel floored at 1e-6) as the training pipeline — see REPRODUCE.md and test_full_parity.js.
This is a 10-class evaluation on clean, single-event recordings. It is not a validation of general Deaf-awareness performance in real homes, where sounds overlap, echo, and arrive through walls. Treat it as evidence the model genuinely learned these ten sounds — nothing more.
We also built and tested a "teach it your own doorbell" feature (few-shot enrollment on the network's embeddings). It failed our own testing: enrolling on 3 door-knock clips, the automatic distance gate false-fired on 36 of 72 non-knock held-out clips, and even a gate tightened to half that distance still false-fired on 15 of 72. A sound-awareness tool that cries wolf is worse than none, so we removed it. The evaluation (eval_knn.py, knn_eval.json) ships in the repository.

04Limits, privacy, honest scope

Not a safety device. SoundCue is an awareness aid and a research prototype. Do not rely on it for emergencies, smoke or CO alarms, medical monitoring, or any situation where a missed sound could hurt someone. It was evaluated on 80 clean test clips of 10 sound classes — not in real homes.
Known limitationWhat it means for you
Clean training dataESC-50 clips are isolated single events. Overlapping sounds, echoes and distance will reduce accuracy. The noise slider in section 02 shows this happening live.
Known confusionsOn the test fold: siren mistaken for car horn (2), and single mix-ups of dog as glass breaking, footsteps as door knock, and footsteps as thunder. Five errors in 80 clips.
4-second window, every 2 sVery short sounds can be diluted inside a window; very long sounds repeat. Repeated identical detections within 10 s merge into one timeline event.
Microphone qualityLaptop and phone mics differ; browser noise suppression can alter the signal. Chrome, Edge or Firefox on HTTPS (or localhost) is required for microphone access. Our own six recorded clips (three knocks, six sirens across two devices) classified as expected, but that is a spot check, not a home validation.

Privacy

Audio is processed on this device only and never leaves it. Nothing is recorded, stored on a server, or uploaded — the page makes no network calls after load. Your event log lives in this browser's local storage; clear it any time. Microphone permission can be revoked in the browser address bar.