---
language:
  - multilingual
tags:
  - speech
  - language-identification
  - ecapa-tdnn
  - onnx
  - coreml
  - kesha-voice-kit
license: apache-2.0
library_name: onnx
pipeline_tag: audio-classification
base_model: speechbrain/lang-id-voxlingua107-ecapa
---

# SpeechBrain ECAPA-TDNN — ONNX + CoreML

Pre-converted [speechbrain/lang-id-voxlingua107-ecapa](https://huggingface.co/speechbrain/lang-id-voxlingua107-ecapa) model for spoken language identification. Supports **107 languages** from audio.

Converted for use with [Kesha Voice Kit](https://github.com/drakulavich/kesha-voice-kit) — open-source voice toolkit.

## Files

| File | Format | Size | Description |
|---|---|---|---|
| `lang-id-ecapa.onnx` | ONNX | ~760KB | Model graph |
| `lang-id-ecapa.onnx.data` | ONNX | ~85MB | Model weights (external data) |
| `lang-id-ecapa.mlpackage.tar.gz` | CoreML | ~40MB | CoreML model archive (macOS) |
| `labels.json` | JSON | <1KB | 107 ISO 639-1 language codes |

## Usage with Kesha Voice Kit

```bash
bun install -g @drakulavich/kesha-voice-kit
kesha install          # downloads this model automatically
kesha --json audio.ogg # transcribe + detect language
```

## Usage with ONNX Runtime (Python)

```python
import onnxruntime as ort
import numpy as np
import json

session = ort.InferenceSession("lang-id-ecapa.onnx")
with open("labels.json") as f:
    labels = json.load(f)

# Input: 16kHz mono float32 waveform
audio = np.random.randn(1, 160000).astype(np.float32)  # 10 seconds
result = session.run(None, {"waveform": audio})
probs = result[0][0]

top_idx = np.argmax(probs)
print(f"Language: {labels[top_idx]} (confidence: {probs[top_idx]:.4f})")
```

## Usage with ONNX Runtime (Rust)

```rust
use ort::session::Session;

let session = Session::builder()?.commit_from_file("lang-id-ecapa.onnx")?;
// Input: "waveform" [1, samples] float32
// Output: "language_probs" [1, 107] float32
```

## Model Details

- **Architecture:** ECAPA-TDNN (originally for speaker recognition, adapted for language ID)
- **Training data:** [VoxLingua107](http://bark.phon.ioc.ee/voxlingua107/) — 6628 hours of speech across 107 languages
- **Input:** Raw waveform at 16kHz mono (`[1, samples]` float32)
- **Output:** Language probabilities (`[1, 107]` float32, softmax applied)
- **Error rate:** 6.7% on VoxLingua107 dev set

## Supported Languages

ab, af, am, ar, as, az, ba, be, bg, bn, bo, br, ca, ceb, cs, cy, da, de, el, en, eo, es, et, eu, fa, fi, fo, fr, gl, gn, gu, ha, haw, he, hi, hr, ht, hu, hy, ia, id, is, it, ja, jw, ka, kk, km, kn, ko, la, lb, ln, lo, lt, lv, mg, mi, mk, ml, mn, mr, ms, mt, my, ne, nl, nn, no, oc, pa, pl, ps, pt, ro, ru, sa, sd, si, sk, sl, sn, so, sq, sr, su, sv, sw, ta, te, tg, th, tk, tl, tr, tt, uk, ur, uz, vi, war, yi, yo, zh

## Conversion

Converted from PyTorch using `torch.onnx.export` (ONNX) and `torch.export` + `coremltools` (CoreML).

Conversion script: [scripts/convert-lang-id-model.py](https://github.com/drakulavich/kesha-voice-kit/blob/main/scripts/convert-lang-id-model.py)

## License

Apache 2.0 (same as the original SpeechBrain model)