Version History¶

This page lists the VTC versions, explains how to migrate from older ones, and compares VTC with LENA.

VTC versions¶

VTC 2.x is built on BabyHuBERT, a self-supervised speech model trained on child-centred audio, which explains most of the accuracy gain over earlier versions. The largest gains over VTC 1.0 are for other children (OCH) and adult males (MAL), both more than 25 F1 points higher in VTC 2.2. Per-class scores are on the home page.

VTC 2.2 uses the VTC 2.1 weights but processes each recording with overlapping sliding windows (75% overlap by default, set by --stride_pct) and averages the predictions of overlapping windows. This improves accuracy for every category and makes processing slower (see How long will it take?). VTC 2.2 is also the first version installed as a Python package that downloads the model from Hugging Face.

Migrating from VTC 2.0 or 2.1¶

• The model is downloaded automatically from Hugging Face on the first run. git-lfs and git clone --recurse-submodules are no longer needed, and you can delete the VTC-2/ folder. After pulling the new version, run uv sync once.

• New commands replace the scripts in scripts/:

  Before	Now
  uv run scripts/infer.py ...	uv run vtc-infer ...
  sh scripts/run.sh	sh run.sh
  uv run scripts/convert.py --input <in> --output <out>	uv run convert-audios --wavs <in> --output <out>

• Renamed options:

  Before	Now
  --high_precision	--high-precision
  --thresholds path/to/custom.toml	--thresholds-location path/to/custom.toml --thresholds-name custom (see Custom threshold files)
  --checkpoint VTC-2/model/best.ckpt, --config VTC-2/model/config.toml	--checkpoint 2.2, --config 2.2 (the defaults), or a local path

• Earlier models: --checkpoint 2.1 loads the same weights as 2.2, but results still differ slightly from VTC 2.1 runs because of the sliding windows. The VTC 2.0 model cannot be selected by version number: it only runs from the v2.0.2 tag of the repository (see Using VTC 2.0).

• The output folder now also contains the threshold file used for the run (f1.toml or hp.toml).

Using VTC 2.0¶

The current version of VTC cannot run the VTC 2.0 model. To use it, for example to reproduce published results, clone the repository at the v2.0.2 tag, the last 2.0 release, in a separate folder. This version needs git-lfs to download the model:

git lfs install
git clone --branch v2.0.2 --recurse-submodules https://github.com/LAAC-LSCP/VTC.git VTC-2.0
cd VTC-2.0
uv sync

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

This version is run with scripts/infer.py and has different options from the current vtc-infer. See the README of the v2.0.2 tag for its options.

Migrating from VTC 1.0¶

• Labels: the CHI label (KCHI and OCH combined) and the SPEECH label no longer exist. To recover CHI, combine KCHI and OCH in your own scripts.
• Output: VTC 2 still writes RTTM files, and also writes a CSV.
• Installation: VTC 2 cannot run in a VTC 1.0 conda environment. Install it from scratch following Getting Started.

VTC vs. LENA¶

VTC is not a drop-in replacement for LENA. Their categories are similar but not equivalent, and the two systems produce different counts on the same audio. Be careful when comparing VTC results with LENA results. For accuracy comparisons, see the ExELang book.