--- license: apache-2.0 base_model: Qwen/Qwen3-ForcedAligner-0.6B base_model_relation: quantized pipeline_tag: automatic-speech-recognition library_name: onnxruntime language: - en - zh - yue - fr - de - it - ja - ko - pt - ru - es tags: - onnx - onnxruntime - qwen3_asr - forced-alignment - word-level - qwen3 --- # Qwen3-ForcedAligner-0.6B (Q4 ONNX) ## Source | Field | Value | |---|---| | Upstream model | [Qwen/Qwen3-ForcedAligner-0.6B](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B) | | ONNX artifact source | [valoomba/Qwen3-ForcedAligner-0.6B-ONNX](https://huggingface.co/valoomba/Qwen3-ForcedAligner-0.6B-ONNX) | | Upstream source revision | `c7cbfc2048c462b0d63a45797104fc9db3ad62b7` | | ONNX artifact source revision | `261c9ed100c1b18a4a1fbc488e05625dc9a4ae5c` | | Packaging source revision | `4b904d4e0eb18c8dd1726ff5e830f72c06dd665b` | | Export tool/script | Community ONNX export (valoomba) repackaged for checksum-verified distribution | | Quantization recipe | Q4 ONNX (`onnx/model_q4.onnx`, self-contained) | ## Model - **Architecture**: Encoder (24L, d=1024) + 28L decoder body, non-autoregressive - **Output**: Word-level timestamps via slot injection; 5000-class softmax x 80ms resolution - **Max audio duration**: 5 minutes - **Inference**: Single forward pass - **Languages**: 11 (English, Chinese, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish) Architecture and capability figures above are inherited from the upstream / valoomba source and are not independently re-measured here. ## Files | File | Size | SHA256 | |---|---|---| | `onnx/model_q4.onnx` | ~1000 MB | `59b528896d70b34e57838e160d16d5f7cfc02d86c7c6ad46cdc57c25c15497b7` | | `config.json` | — | — | | `tokenizer.json` | — | — | | `vocab.json` | — | — | | `merges.txt` | — | — | | `preprocessor_config.json` | — | — | | `tokenizer_config.json` | — | — | | `special_tokens_map.json` | — | — | | `added_tokens.json` | — | — | Note: `model_q4.onnx` is self-contained (no external data sidecar required). ## Intended Use Word-level forced alignment for local inference via ONNX Runtime. ## Runtime Notes - Designed for ONNX Runtime compatible runtimes. - Input: 16 kHz mono audio. - Inference: single forward pass, non-autoregressive. - Q4 quantization: self-contained ONNX with no external sidecar. ## 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 - 5-minute maximum audio duration per inference call. - Q4 quantization may affect timestamp precision at the margin; validate on representative audio before production use. - 11-language coverage; quality by language not benchmarked in this repo. ## License [Apache 2.0](https://choosealicense.com/licenses/apache-2.0/) — see the [original model card](https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B). This packaging repo adds no new license terms.
