indicconformer-assamese-onnx-int8

IndicConformerASR Assamese ONNX INT8

This repository contains an INT8 ONNX conversion of AI4Bharat's IndicConformer Assamese ASR model, prepared by the Assam AI Initiative (AAII) for lightweight speech-recognition deployment, including browser-based voice-typing applications.

The goal of this release is to make IndicConformer easier to run in resource-constrained and client-side environments without requiring the full NVIDIA NeMo stack during inference.

Model

• Base model: AI4Bharat IndicConformer
• Original code: AI4Bharat/IndicConformerASR
• ONNX conversion code: aaiiaxomi/IndicConformerASR_ONNX_Conversion

Conversion

The original NeMo IndicConformer model was exported to ONNX and quantized to INT8 for efficient inference.

The conversion supports an end-to-end ONNX graph in which the speech frontend is included inside the model:

Raw audio
   ↓
STFT / Mel preprocessing
   ↓
Conformer encoder
   ↓
CTC head
   ↓
ASR output

This reduces external preprocessing dependencies and is useful for deployment through ONNX Runtime and browser-oriented inference pipelines.

Files

model_cpu_int8.onnx   # INT8 ONNX ASR model
model_config.json     # model and preprocessing configuration
config.json           # deployment configuration
id2token.json         # token ID → token mapping
tokenizer/            # tokenizer resources

Intended Use

This model was converted primarily for AAII's web-based voice-typing and language-technology applications, with emphasis on efficient client-side or CPU inference.

Possible uses include:

• browser-based voice typing
• lightweight speech-to-text applications
• educational and accessibility tools
• prototyping ASR applications for Indian languages

Limitations

• This is a quantized deployment version of the original model; small numerical or accuracy differences may occur compared with the original NeMo checkpoint.
• The exported deployment path uses CTC decoding.
• Performance depends on the language, recording conditions, speaker characteristics, and domain.
• Users should evaluate the model on their target data before production use.

Acknowledgements

This model is derived from AI4Bharat's IndicConformer, a suite of ASR models for Indian languages.

We gratefully acknowledge the AI4Bharat team for developing and releasing IndicConformer and its associated resources:

• AI4Bharat IndicConformer — Hugging Face
• AI4Bharat IndicConformerASR — GitHub

The ONNX export, preprocessing integration, optimization, and INT8 conversion utilities used for this release are available here:

• AAII IndicConformer ONNX Conversion

Please cite and acknowledge the original AI4Bharat IndicConformer work/model release when using this converted model in research or downstream applications.

License

The original ai4bharat/IndicConformer Hugging Face model is released under CC-BY-4.0. This converted model should be used in accordance with the terms of the original model release.

Please also check the upstream model card and repository for the latest licensing and usage information.