Threshold selection¶

VTC has a tunable "sensitivity" setting for each speaker category. The default settings work well for most users; this page is for those who want to deliberately trade off catching every segment ("recall") against avoiding wrong labels ("precision").

What thresholds actually do¶

For every short slice of audio (a few hundredths of a second), VTC produces four numbers between 0 and 1, one per speaker category, representing how confident the model is that the category is currently speaking. To turn those continuous scores into yes/no decisions ("Was the key child speaking at this moment? Yes or no?"), VTC compares each score against a threshold. If the score is above the threshold, the category is marked as active for that slice; otherwise it is not.

You can think of the threshold as the bar a category has to clear before VTC will commit to a label:

Lower threshold → easier to clear → VTC labels more segments. Good if you don't want to miss anything (high recall), but you'll get more wrong labels too.
Higher threshold → harder to clear → VTC labels fewer segments, but the ones it does label are more likely to be correct (high precision).

There is no universally best threshold, the right value depends on whether you care more about avoiding misses (favouring lower thresholds) or avoiding false alarms (favouring higher thresholds).

Built-in threshold sets¶

VTC ships with two pre-tuned threshold sets. The default balances recall and precision; the high-precision set favours fewer wrong labels at the cost of missing more speech.

Threshold set	File	When to use it
Balanced (default)	`thresholds/f1.toml`	General use — maximises overall F1 (a balance of recall and precision).
High precision	`thresholds/hp.toml`	When you would rather miss some speech than risk introducing false detections.

Both threshold sets were tuned on the BabyTrain-2025 validation set. The balanced set is selected automatically; to use the high-precision set, add the --high_precision flag, or pass the file directly with --thresholds:

uv run scripts/infer.py      \
    --wavs <audio_folder>    \
    --output <output_folder> \
    --high_precision \
    --device cpu

Try it interactively¶

The widget below lets you move each threshold up and down and watch the effect on overall accuracy in real time. This is the easiest way to build intuition for what thresholds do.

Drag each slider, or click directly on a chart, to set the decision threshold for each category. The macro-F1 number at the top updates live as you tune.

What the curves show

Each curve traces how the F1 score for one category changes as you sweep its threshold from 0 to 1. The peak of each curve is the value that maximises that category's F1 in isolation; the default f1.toml thresholds correspond to those peaks.

Technical Reference¶

Implementation details for engineers tuning their own thresholds.

Custom threshold files¶

To use your own thresholds, create a file custom.toml following the TOML specification and pass it to the inference script with --thresholds:

uv run scripts/infer.py      \
    --wavs <audio_folder>    \
    --output <output_folder> \
    --thresholds thresholds/custom.toml \
    --device cpu

Threshold tuning workflow¶

To re-tune thresholds on your own annotated data:

Run inference on your validation set with --save_logits to dump per-frame model scores.
Run the tuning pipeline (a per-class grid search) against your reference RTTM annotations.
The pipeline emits the per-class thresholds maximising your chosen objective (default: macro-F1).

Per-class independence

Each class's threshold is optimised independently — the search does not jointly account for cross-class trade-offs.