polvallverdu/voxlingua107-onnx

VoxLingua107 ECAPA-TDNN language ID (ONNX mirror)

A pinned, unmodified mirror of christopherthompson81/voxlingua107-lid-onnx (revision e02e1da805ae49635fe1aa7913c3f1e7f5f5fde6), an ONNX export of speechbrain/lang-id-voxlingua107-ecapa. It is the default model source of the voxlingua107 npm package, which falls back to the original repository if this one is unavailable.

Nothing was retrained, converted or modified.

File	SHA-256
voxlingua107.onnx	e2c3c3da39b99e3f9196d15fceef6a65f702320038bbc08813a4f21280255ce8
lang_map.json	d593825e7e8ce82c8b50c0810a371a3e69cd0293b52751ddf515f6ee2c592182

manifest.json is the original upstream manifest (MD5 hashes).

Model

• Input audio: float32 [1, samples], 16 kHz mono, values in [-1, 1]. Feature extraction (fbank + mean normalization) is inside the graph.
• Outputs logits: float32 [1, 107] log-probabilities in lang_map.json order; embedding: float32 [1, 256].

The npm package's test suite checks this file against the original PyTorch model: probabilities agree within 6e-6 on the upstream example clip and on synthetic noise, and the top-1 language matches.

Usage (Node.js / Bun)

import { LanguageIdentifier } from "voxlingua107";

const lid = await LanguageIdentifier.load();
const { code, probability } = await lid.identify(samples16k);

Credits and license

Apache-2.0, inherited from the upstream model.

• Model: Jörgen Valk and Tanel Alumäe, trained with SpeechBrain.
• ONNX export: Chris Thompson, see the conversion scripts.

@inproceedings{valk2021slt,
  title={{VoxLingua107}: a Dataset for Spoken Language Recognition},
  author={J{\"o}rgen Valk and Tanel Alum{\"a}e},
  booktitle={Proc. IEEE SLT Workshop},
  year={2021},
}