---
language: en
library_name: transformers.js
pipeline_tag: automatic-speech-recognition
license: apache-2.0
base_model: vitouphy/wav2vec2-xls-r-300m-timit-phoneme
tags:
  - phoneme-recognition
  - pronunciation
  - onnx
---

# Speako Phoneme Recognizer

ONNX export of [vitouphy/wav2vec2-xls-r-300m-timit-phoneme](https://huggingface.co/vitouphy/wav2vec2-xls-r-300m-timit-phoneme) (Apache-2.0) for in-browser phoneme recognition with [Transformers.js](https://huggingface.co/docs/transformers.js). Used by [Speako](https://speako.tre.systems/) for pronunciation feedback: recognized IPA phonemes are aligned against CMUdict reference pronunciations to score each word.

- `onnx/model_quantized.onnx` (~355 MB): INT8 dynamic quantization of MatMul ops only — quantized Conv layers crash onnxruntime, so the convolutional feature extractor stays fp32.
- `tokenizer.json` is synthesized from `vocab.json` (the source repo ships only a slow CTC tokenizer, which Transformers.js cannot load).
- Output: IPA phonemes from a 39-symbol TIMIT-derived inventory, words separated by spaces. Run on CPU/WASM.

```js
import { pipeline } from '@huggingface/transformers';
const asr = await pipeline('automatic-speech-recognition', 'robg/speako-phoneme-recognizer', {
  dtype: 'q8',
  device: 'wasm',
});
const { text } = await asr(float32Audio16kHz); // "ɪn ʤɛnɝəl tɛknɑləʤi ..."
```