Skip to content

voxlingua107_export

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

voxlingua107_export

README.md

VoxLingua107 Language-ID Export

Exports speechbrain/lang-id-voxlingua107-ecapa into a single end-to-end ONNX graph for use as the language-identification backend in Vernacula.

The model is an ECAPA-TDNN classifier over 80-dim mel filterbank features, trained on the VoxLingua107 corpus. It takes raw 16 kHz mono audio and emits both a 107-class logit vector (for language classification) and a 192-dim pooled embedding (useful for language-similarity clustering). License is Apache 2.0, so it can be bundled or auto-downloaded without a gated-accept flow.

Files

  • export_voxlingua_to_onnx.py — exports voxlingua107.onnx + lang_map.json
  • verify_voxlingua_parity.py — PyTorch ↔ ONNX parity check on real audio clips
  • src/ecapa_wrapper.py — thin nn.Module composing SpeechBrain's submodules for a clean export graph
  • src/lang_map.py — builds the {index: {iso, name}} lookup that ships alongside the ONNX
  • requirements.txt — Python dependencies for the export environment

Output graph

Single file, no external data.

Tensor Shape Dtype Notes
audio [batch, samples] float32 16 kHz mono, variable length
logits [batch, 107] float32 Language classification logits
embedding [batch, 256] float32 Pooled ECAPA embedding

Preprocessing (FBANK, per-utterance mean-variance normalisation) is folded into the graph, so the C# runtime just sends raw PCM.

Environment

Use Python 3.11 or 3.12.

python3 -m venv .venv-voxlingua-export
source .venv-voxlingua-export/bin/activate
pip install -r public/scripts/voxlingua107_export/requirements.txt

Export

python public/scripts/voxlingua107_export/export_voxlingua_to_onnx.py \
    --out-dir ./voxlingua107

This produces:

voxlingua107/
├── voxlingua107.onnx     # ~83 MB FP32, weights inlined
└── lang_map.json         # 107 entries

PyTorch's dynamo-based exporter (torch>=2.8) splits weights into a companion .onnx.data file by default. The script re-saves the graph inline once export finishes — this model is small enough (well under the 2 GB protobuf ceiling) that the sidecar just adds distribution friction.

The script downloads the SpeechBrain checkpoint into ./.voxlingua107-cache/ on first run; subsequent runs reuse the cached weights.

Parity verification

Prepare a handful of 16 kHz mono WAV clips across several languages (ten clips spanning en, zh, hi, ar, es, de, fr, ru, ja, sw is a good baseline), then:

python public/scripts/voxlingua107_export/verify_voxlingua_parity.py \
    --model-dir ./voxlingua107 \
    --clips clip_en.wav clip_zh.wav clip_hi.wav ...

Pass criteria:

  • Top-1 language matches between PyTorch and ONNX on every clip
  • Softmax probability max-abs-diff ≤ 1e-3 (scale-invariant; raw-logit diff grows with clip length because of FP32 accumulation, so we compare probs rather than logits)
  • Embedding cosine similarity ≥ 0.9999

Observed values on a 5-clip test set (en, de, fr, ru, hu, durations 90 s– 602 s) are Δprob 3e-11 to 6e-5 and cosine 1.000000 across the board, so the gates have comfortable headroom.

Failures print the reason(s) and return a non-zero exit code so the check can gate CI if needed.

Publishing

After a successful export, build the manifest and upload to christopherthompson81/voxlingua107-lid-onnx using the shared tools at the scripts/ root:

python scripts/make_manifest.py \
    --model-dir ~/models/voxlingua107 \
    --files voxlingua107.onnx lang_map.json

python scripts/upload_to_hf.py \
    --model-dir ~/models/voxlingua107 \
    --repo-id christopherthompson81/voxlingua107-lid-onnx \
    --sync-readme

The model card is sourced from scripts/hf_readmes/voxlingua107-lid-onnx/README.md; edit it there, commit, and re-run with --sync-readme to update HF.