Threshold selection¶

Each speaker category has a threshold that controls how readily VTC assigns it. The default thresholds suit most uses. Change them if you want to favour finding all speech (recall) or avoiding wrong labels (precision).

What thresholds do¶

For every frame of audio (a few hundredths of a second), the model outputs four scores between 0 and 1, one per category, that express how likely it is that this category is speaking. VTC marks a category as active in that frame when its score is above the category's threshold.

A lower threshold produces more segments: less speech is missed (higher recall), but more labels are wrong.
A higher threshold produces fewer segments, but more of them are correct (higher precision).

The right choice depends on whether missed speech or false detections matter more for your analysis.

Built-in threshold sets¶

Threshold set	Name	When to use it
Balanced (default)	`f1`	General use. Maximises F1, which balances recall and precision.
High precision	`hp`	When you would rather miss some speech than get false detections.

Both sets were tuned on the BabyTrain-2025 validation set. They are downloaded with the model, from the thresholds/ folder of the coml/VTC-2 Hugging Face repository. To use the high-precision set, add --high-precision:

uv run vtc-infer             \
    --wavs <audio_folder>    \
    --output <output_folder> \
    --high-precision         \
    --device cpu

VTC copies the threshold file it used (f1.toml or hp.toml) into the output folder.

Try it interactively¶

Drag the sliders, or click on a chart, to set the threshold of each category. The macro-F1 at the top updates as you go.

What the curves show

Each curve shows a category's F1 score as its threshold goes from 0 to 1. The default f1.toml thresholds are at the peaks of these curves.

Technical Reference¶

Custom threshold files¶

Write your thresholds in a TOML file, such as my_thresholds/custom.toml, with one section per class. The values below are those of the default f1 set:

[KCHI]
lower_bound = 0.437
upper_bound = 1.0

[OCH]
lower_bound = 0.378
upper_bound = 1.0

[MAL]
lower_bound = 0.429
upper_bound = 1.0

[FEM]
lower_bound = 0.377
upper_bound = 1.0

Pass the file's location with --thresholds-location and its name, without .toml, with --thresholds-name:

uv run vtc-infer             \
    --wavs <audio_folder>    \
    --output <output_folder> \
    --thresholds-location my_thresholds/ \
    --thresholds-name custom \
    --device cpu

--thresholds-location accepts the .toml file itself or a folder. A folder is searched recursively and must contain exactly one file with that name. Without --thresholds-location, VTC looks in the same place as the model given by --checkpoint. --high-precision overrides --thresholds-name, so it looks for an hp.toml file in --thresholds-location. The thresholds file is copied into the output folder under its own name, here custom.toml.

Threshold tuning workflow¶

To re-tune thresholds on your own annotated data:

Run inference on your validation set with --save_probs to save the per-frame scores.
Run the tuning pipeline, a per-class grid search, against your reference RTTM annotations.
The pipeline outputs the per-class thresholds that maximise the chosen objective (macro-F1 by default).

Per-class independence

Each class's threshold is optimised separately; the search does not account for trade-offs between classes.