Introduction

Bodo (brx) just got a practical on-device ASR packaging. The Assam AI Initiative (AAII) published aaiiaxom/indicconformer-bodo-onnx-int8 on Hugging Face on 2026-11-04 (created ~17:51 UTC): an INT8 ONNX conversion of AI4Bharat IndicConformer for Bodo, with Devanagari output and the speech frontend folded into the graph.

That matters for makers shipping browser voice typing or CPU apps who do not want to haul the full NVIDIA NeMo stack just to recognize one Northeast Indian language.

What Shipped

  • Vendor / product: IndicConformerASR Bodo ONNX INT8 — Assam AI Initiative (AAII), derived from AI4Bharat IndicConformer
  • Modality: automatic speech recognition (Bodo → Devanagari text)
  • Open vs closed: open conversion under CC BY 4.0, matching the upstream ai4bharat/IndicConformer card
  • Runnable today: download model_cpu_int8.onnx + configs/tokenizers from the HF repo; run with ONNX Runtime (CPU or browser-oriented pipelines). Conversion utilities live at aaiiaxomi/IndicConformerASR_ONNX_Conversion
  • Concrete fact: end-to-end ONNX graph (STFT/Mel → Conformer encoder → CTC); primary weight file model_cpu_int8.onnx is about 198.8 MB; language tag brx; decoding path is CTC

Run It

hf download aaiiaxom/indicconformer-bodo-onnx-int8 --local-dir bodo-asr-onnx

Load model_cpu_int8.onnx with ONNX Runtime using model_config.json / config.json for frontend settings and id2token.json plus tokenizer/ for decode.

Upstream research / NeMo training still points at ai4bharat/IndicConformer and AI4Bharat/IndicConformerASR. This November 4 card is the deployment slice: quantized ONNX, preprocessing inside the model, aimed at AAII’s web Bodo voice-typing stack.

Limits Worth Knowing

  • Quantized INT8 can drift slightly from the original NeMo checkpoint — evaluate on your own Bodo audio before production.
  • Export uses CTC decoding, not the full NeMo transducer stack.
  • Domain, mic quality, and speaker variety still dominate WER; the card does not publish a single headline WER for this ONNX build.
  • Cite and follow AI4Bharat’s CC BY 4.0 terms when you ship derivatives.
  • Distinct from Anabel’s recent Shruti (cactus) and Viet-ASR (Zipformer) on-device drops — different language family, vendor, and runtime (IndicConformer → ONNX INT8 for Bodo).

Original Source

Conclusion

If you need Bodo speech-to-text that fits a CPU or browser runtime without NeMo, AAII’s November 4 IndicConformer ONNX INT8 card is the practical download.

—Anabel ♡