iTantra STT: built INT8 models
Quantized ONNX models used by the iTantra STT engine and its in-browser demo (https://huggingface.co/spaces/zxnjxd/itantra-stt-browser). Code: https://github.com/ (iTantra STT repository).
| Folder | Built from | Licence |
|---|---|---|
| indic-conformer-int8 | ai4bharat/indic-conformer-600m-multilingual, per-channel INT8, depthwise convs FP32 | MIT (© AI4Bharat) |
| indic-conformer-int4mixed | same model; FFN MatMuls of layers 2–23 block-wise INT4 (MatMulNBits), the rest as INT8 | MIT (© AI4Bharat) |
| lid-ambernet | NVIDIA NeMo langid_ambernet → ONNX, static INT8 | NVIDIA NGC Terms of Use |
| whisper-tiny-en | onnx-community/whisper-tiny.en (INT8) | MIT (© OpenAI) |
| silero-vad | onnx-community/silero-vad | MIT |
| mt-* | ONNX INT8 exports (hari31416) of AI4Bharat IndicTrans2 distilled models | MIT (© AI4Bharat) |
Each model keeps its upstream licence; using these files means accepting those terms. AI4Bharat asks users of IndicConformer to share contact details on its upstream page: if you build on it, please also request access there.