scribe-ecapa-voxlingua107

VoxLingua107 ECAPA-TDNN (ONNX) for Scribe SV

A byte-for-byte copy of christopherthompson81/voxlingua107-lid-onnx, published here so that Scribe SV — a Windows dictation and translation utility — can download it on demand from a repository we control. Nothing was retrained, converted or modified; the voxlingua107.onnx file has the same MD5 (90101ebfee08d96afe7d516e988910fa) as the upstream one.

What Scribe SV uses it for

On machines without a CUDA GPU, dictation runs on Parakeet TDT 0.6b v3, which covers 25 European languages. Speech in any other language would come back as plausible-looking nonsense (we measured 106.9 % WER on Turkish), so before recognising an audio file we ask this model one binary question: is the language of this audio inside the Parakeet tier, or outside it?

The answer is obtained by masking the 107-class softmax down to the 25 Parakeet languages and comparing the largest remaining probability against a threshold. The predicted language name is never used — inside the tier Parakeet writes in the language it hears, and outside it Whisper detects the language itself. Only if the verdict is "outside" does the application offer to download crash-sv/scribe-whisper-turbo-int8.

Measured on our bench

FLEURS clips, CPU (CPUExecutionProvider), 10-second segments, 200 clips per language:

Errors inside the mask (e.g. Russian scored as Ukrainian) are harmless for us: both languages are in the tier, and the name is discarded anyway.

Contents

Preprocessing (FBANK, per-utterance mean-variance normalisation) is folded into the graph, so callers send raw PCM and nothing else.

Usage

import json
import numpy as np
import onnxruntime as ort
from huggingface_hub import snapshot_download

folder = snapshot_download("crash-sv/scribe-ecapa-voxlingua107")
session = ort.InferenceSession(f"{folder}/voxlingua107.onnx",
                               providers=["CPUExecutionProvider"])
(logits,) = session.run(["logits"],
                        {session.get_inputs()[0].name: pcm_16k_float32[None, :]})

Credits

• ONNX export: christopherthompson81/voxlingua107-lid-onnx (part of Vernacula) — the STFT op is replaced by two Conv1D passes so the whole preprocessing path has CUDA kernels.
• Upstream model: speechbrain/lang-id-voxlingua107-ecapa by the SpeechBrain team, trained on the VoxLingua107 dataset (CC-BY-4.0).

Citation

Jörgen Valk and Tanel Alumäe. VoxLingua107: a Dataset for Spoken Language Recognition. Proc. IEEE SLT Workshop, 2021.

<!-- legal:begin — generated by tools/legal_package.py, do not edit -->

Attribution and license

ECAPA-TDNN VoxLingua107 spoken language ID (ONNX)

• Author: SpeechBrain; Jörgen Valk, Tanel Alumäe (VoxLingua107); Christopher Thompson (ONNX export)
• License: Apache-2.0 — https://www.apache.org/licenses/LICENSE-2.0
• Upstream: https://huggingface.co/christopherthompson81/voxlingua107-lid-onnx
• Upstream: https://huggingface.co/speechbrain/lang-id-voxlingua107-ecapa
• Changes: Unchanged copy of christopherthompson81/voxlingua107-lid-onnx (ONNX export of speechbrain/lang-id-voxlingua107-ecapa).

License text (LICENSE-Apache-2.0.txt) lies in the repository root.

This republication keeps the original licenses.

<!-- legal:end -->