Wav2vec2 Lv60 Espeak Ft ONNX

--- license: apache-2.0 base_model: facebook/wav2vec2-lv-60-espeak-cv-ft base_model_relation: quantized pipeline_tag: automatic-speech-recognition library_name: onnxruntime language: - multilingual tags: - onnx - onnxruntime - wav2vec2 - forced-alignment - phoneme --- # wav2vec2-lv-60-espeak-cv-ft (INT8 ONNX) ## Source | Field | Value | |---|---| | Upstream model | [facebook/wav2vec2-lv-60-espeak-cv-ft](https://huggingface.co/facebook/wav2vec2-lv-60-espeak-cv-ft) | | ONNX artifact source | [onnx-community/wav2vec2-lv-60-espeak-cv-ft-ONNX](https://huggingface.co/onnx-community/wav2vec2-lv-60-espeak-cv-ft-ONNX) | | Upstream source revision | `ae45363bf3413b374fecd9dc8bc1df0e24c3b7f4` | | ONNX artifact source revision | `c69750f5043e5e1f8a71ab95dd3b98338c280c92` | | Export tool/script | Mirrored INT8 ONNX from onnx-community | | Quantization recipe | INT8 ONNX (`onnx/model_int8.onnx`) | ## Model - **Architecture**: Wav2Vec2-Large-LV60, fine-tuned on multilingual CommonVoice for IPA phoneme recognition - **Output**: CTC logits over 392 IPA phoneme classes at 20ms/frame (50 fps) - **Sample rate**: 16 kHz input required - **Languages**: 50+ via zero-shot cross-lingual transfer ## Files | File | Size | SHA256 | |---|---|---| | `onnx/model_int8.onnx` | ~303 MB | `74174710e34035bbb7f611601d016c32fc575de7a6f53b1078107dc10a84e7ae` | | `vocab.json` | — | — | | `config.json` | — | — | | `preprocessor_config.json` | — | — | | `tokenizer_config.json` | — | — | ## Intended Use CTC phoneme recognition / forced-alignment backbone for local inference via ONNX Runtime. ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Input: 16 kHz mono audio. - Output: CTC phoneme logits per frame (50 fps / 20ms resolution). - INT8 quantization; validate on the target execution provider before production use. ## Precision and Packaging Export tooling, precision, and quantization are recorded in the **Source** table above. This packaging mirror does not publish independent parity benchmarks; validate on your target execution provider before production use. ## Limitations - Requires 16 kHz mono input; other sample rates must be resampled. - Zero-shot multilingual transfer; phoneme accuracy varies by language and script. Evaluate on target languages before production use. - INT8 quantization may affect phoneme boundary precision; validate on representative audio. ## License [Apache 2.0](https://choosealicense.com/licenses/apache-2.0/) — see the [original model card](https://huggingface.co/facebook/wav2vec2-lv-60-espeak-cv-ft). This packaging repo adds no new license terms.