Downloads · 30 days
9
5% of all-time downloads
indic-conformer-600m-int8-onnx
livinNector/indic-conformer-600m-int8-onnx
indic-conformer-600m-int8-onnx is a machine learning model from livinNector. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Int8 dynamically-quantized copy of ai4bharat/indic-conformer-600m-multilingual (MIT): the ONNX encoder is quantized from 2.4 GB fp32 (external data) to a self-contained 0.62 GB int8 file; everything else (CTC decoder,…
All-time downloads
197
Public
Repo size
772 MB
Likes
1
Public
Repo makeup
Click a slice to open those files.
.onnx735 MB · 95%
At a glance
- License
- mit
- Access
- Public
- Created
- Aug 5, 2026
- Updated
- Aug 5, 2026
- SHA
- a5eec51d
Base models
Model card
From the Hugging Face model README
indic-conformer-600m-int8-onnx
Int8 dynamically-quantized copy of ai4bharat/indic-conformer-600m-multilingual (MIT): the ONNX encoder is quantized from 2.4 GB fp32 (external data) to a self-contained 0.62 GB int8 file; everything else (CTC decoder, per-language heads, preprocessor, remote code) is unchanged.
Why: the fp32 encoder needs ~2.6 GB resident — too much next to an LLM on an 8 GB device. Int8 cuts peak RSS to ~1.1 GB and speeds up loading ~3×, with a measured WER cost of ≈ +0.5 points on one of six spot-checked conversational clips (the rest bit-identical).
from transformers import AutoModel
model = AutoModel.from_pretrained("yashwork-byte/indic-conformer-600m-int8-onnx",
trust_remote_code=True)
text = model(wav_16k_mono_tensor, "hi", "ctc")
Built for the voice-note-digester
project (quantize_stt.py — onnxruntime dynamic quantization, QInt8
weights). All credit for the model to AI4Bharat;
this repo only repackages it for memory-constrained on-device use.
- License
- mit
- Languages
- hi, bn, ta, as, gu, kn
- Created
- Aug 5, 2026
- Updated
- Aug 5, 2026