indic-conformer-600m-multilingual

An unmodified copy of AI4Bharat's ai4bharat/indic-conformer-600m-multilingual, a 600M-parameter hybrid CTC + RNNT Conformer for the 22 official languages of India. Mirrored by Diction Labs so Diction's speech-to-text services pull from a repository we control.

Every weight and code file is byte-identical to the upstream repository (LFS sha256 compared file by file on 2026-10-11). Only this README and the LICENSE file were added.

Credit

Model, training and code: AI4Bharat (IIT Madras). Upstream repository: https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual. Please cite AI4Bharat's IndicVoices / IndicConformer work if you use it.

Licence

MIT, as released upstream. See LICENSE.

Use

Same as upstream:

from transformers import AutoModel
import torch, torchaudio

model = AutoModel.from_pretrained("DictionLabs/indic-conformer-600m-multilingual", trust_remote_code=True)
wav, sr = torchaudio.load("audio.wav")            # 16 kHz mono
print(model(wav, "hi", "ctc"))                    # or "rnnt"

Diction's server uses CTC decoding for Hindi and cuts audio into ~15 s windows: in our tests the model dropped words on continuous speech longer than about 18 s.