Snoring Detection in Sleep Environments

Hypothesis

Snoring could be identified as relatively loud acoustic events.

Methodology

Implemented a simple threshold-based detector. The app sampled the recording's amplitude every 200 ms and treated anything above a fixed threshold as a potential snoring event.

Sounds that came back within 10 seconds were merged into the same event, so that one snore was not counted many times.

Results

In use, the detector fired on many environmental sounds — clocks, conversations and birds — because they also crossed the threshold. Measured on real recordings, it failed in both directions:

• False positives: three of four recordings of clinking sounds were logged as snoring.
• Missed snoring: in a quieter 15-minute stretch of a real night, the threshold fired once — on a sound that was not snoring — and missed every snore below it.
• Merged snoring: breaths are 3–5 s apart, always inside the 10-second merge window, so continuous snoring collapsed into a single event and the night scored 95, "Great".

Limitations

A single threshold measures how loud the room is, not what made the sound. It cannot adapt when the background changes, and it has no notion of a snore's shape.

Takeaway

Amplitude alone is insufficient for distinguishing snoring from environmental noise.

The next iteration focused on the acoustic characteristics of snoring rather than signal intensity.

Hypothesis

A convolutional neural network could learn the distinctive time-frequency patterns of snoring and classify them more accurately than amplitude thresholding.

Methodology

1. Dataset

Used a public snoring dataset from Kaggle (Khan, 2019): one-second clips, split 70 / 15 / 15 into training, validation and test sets.

2. Feature extraction

Converted each clip into a 64-band mel spectrogram. The pipeline assumed every recording was 48 kHz — an assumption later found to be wrong.

3. Model architecture

Designed a lightweight CNN for eventual on-device use. Six channel configurations were compared, prioritising low computational cost.

4. Data augmentation

Each of the 700 training clips gained four augmented copies. The final training set held 3,720 clips: these 3,500, plus 22 one-second segments of my own clinking recordings (as non-snoring) and matching extra snoring copies, each with their augmentations. Augmentation was applied only after the split, so no test clip had a copy in training.

Results

99.3%

initial test accuracy (149 of 150)

98.2%

mean test accuracy across 30 training runs

1 / 900 s

classified as snoring in a real 15-minute night

On a real 15-minute sleep recording, the model labelled only one of 900 seconds as snoring. In the same segment, the final detector (V3) found 32 snoring events; these have not yet been confirmed by manual annotation.

This discrepancy motivated a deeper investigation into the training and evaluation pipeline.

Failure Analysis

1. Spurious correlation — sampling-rate bias

The dataset mixes 44.1 kHz and 48 kHz recordings, and the pipeline treated all of them as 48 kHz without resampling. A 44.1 kHz clip therefore ended in 0.08 s of padded silence and had its frequencies read about 9% too high. More importantly, the sampling rate was strongly correlated with the label.

A rule using only the sampling rate would score about 89% on the full dataset. The CNN might have learned this preprocessing artifact rather than relying only on acoustic characteristics.

Controlled experiment

All 150 test clips were properly resampled to 48 kHz and evaluated with the already-trained model.

The model is sensitive to preprocessing artifacts. The 68.7% is not the expected performance of a model retrained on a corrected pipeline.

2. Model selection bias

The test set was used to choose the final architecture among six candidates, which compromised the independence of the final test score.

A checkpoint bug compounded this: a shallow copy of the model's state dictionary kept changing as training continued, so the "best" snapshot was not preserved. Some earlier figures — including a 100% result — cannot be treated as independent estimates of generalisation.

3. Real-world generalisation

The CNN continued to perform poorly on real sleep recordings. Differences in background noise, distance to the microphone, sound intensity and room acoustics plausibly contributed, but the experiments did not isolate domain shift from the preprocessing and evaluation problems.

Takeaway

High benchmark accuracy does not guarantee real-world performance.

Unintended correlations, inconsistent preprocessing and evaluation design can significantly distort reported performance. This motivated an approach built on domain-specific acoustic characteristics and temporal context.

Hypothesis

Snoring is characterised not only by its acoustic properties but also by its temporal relationship to breathing: snores recur with each breath, while most bedroom noises happen once. Requiring both a snore-like sound and a breathing-like rhythm should therefore separate snoring from one-off noises.

Methodology

A lightweight, two-stage rule-based detector that runs on the phone. Stage 1 finds sounds with snore-like acoustics; Stage 2 checks whether they repeat like breathing.

Stage 1 — Acoustic feature analysis

Adaptive noise floor. Instead of a fixed threshold, the background level is estimated from the quietest moments (15th percentile) of a rolling 90-second window, so the detection threshold follows changes in the room.

Feature extraction. Six characteristics from FFT-based spectral analysis and timing:

Snore score = Σ (feature score × weight)

Events scoring at least 0.55 pass Stage 1. Two hard rules reject sounds with excessive high-frequency energy or a noise-like flat spectrum. A sweep over 0.30–0.95 on nine real recordings showed 0.55 to be the strictest value that still keeps every snoring recording.

Stage 2 — Temporal pattern validation

Stage 1 judges each sound alone, and non-snoring sounds can share its acoustics. Stage 2 examines the timing between candidates.

CV = standard deviation of intervals / mean interval

A lower CV means more consistent repetition. Four events are required because three unrelated sounds can land at even spacing by chance; four means three intervals must agree.

Results

Stage 1 — Public dataset

On the 1,000 Kaggle clips, Stage 1 alone reached 80.0% accuracy (78.7% on the CNN's 150-clip test split) without any supervised training. Most misses came from clips filled with sound from start to end, which leave the adaptive floor no quiet moment to measure against.

Full system — Real recordings

The Kaggle clips are isolated one-second recordings, so they cannot test Stage 2, which needs a sequence of events. The complete system was therefore evaluated on real recordings.

These results are promising, but the night has not been manually annotated, so precision and recall on real recordings are not yet measured. The detector costs about 0.039% of one CPU core — roughly 11 seconds of compute for an eight-hour night (measured on a laptop, not yet on a phone).

Limitations

1. Isolated snoring events. At least four consecutive events are required, so isolated or highly irregular snoring may be missed — acceptable for the current goal of finding sustained episodes rather than every single snore.
2. Periodic environmental noise. Regular sounds such as a ticking clock could still satisfy the temporal criteria; none were among the test recordings.
3. Quantitative validation. Manually annotated real-world recordings are needed to measure precision, recall and F1-score. The weights were set by hand from a few recordings.

Takeaway

Temporal context provides information that individual acoustic features cannot.

Combining frequency-domain analysis with temporal validation identified repetitive snoring in real recordings without a trained neural network.