Hima

A small open model that listens: speech in, text out, for 11 Indian languages and English. Training is still running, and a new checkpoint appears here every 10,000 steps.

Try the newest checkpoint in three lines

git clone https://github.com/Prahaladha-Reddy/hima && cd hima
pip install torch torchaudio "transformers>=5.13" soundfile huggingface_hub
python scripts/transcribe.py my_clip.wav

It downloads the newest full/step_N folder (about 2.4 GB) on the first run and prints the text. No login is needed. A clip of 1 to 30 seconds works best. Add --step 140000 to choose a checkpoint.

Which file is which

File	What it is
config.json	names the newest checkpoint. The script reads it first.
full/step_N/model.pt	all the weights at step N. The highest N is the newest.
full/step_N/config.json	the settings of the run
full/step_N/pieces.json	the table of word pieces the model uses
full/step_N/scores.json	the score on a held-out test, by language (share of letters wrong)
full/step_N/commit.txt	the commit of the code repo that made it

What is inside

Part	Job	Starts from
Ears	sound to frames	the audio part of Qwen3-ASR-1.7B
Bridge	fit the frames to the writer	new, small
Writer	write the text, piece by piece	Gemma 3 270M
Vote layer	a helper that is used only while training	new, small

Code and notes: https://github.com/Prahaladha-Reddy/hima

Limits

• A research checkpoint, not a product. No result is claimed yet.
• Scores differ a lot between languages; read scores.json.
• The model is not told which language it hears. It sometimes writes a clip in the letters of another language.
• Clips longer than 30 seconds were not used in training.

Licence

• Ours: Apache 2.0. The code, the bridge, the vote layer and the changes that training made (see LICENSE).
• The ears start from Qwen3-ASR-1.7B, which is Apache 2.0.
• The writer starts from Gemma 3 270M. That part stays under the Gemma Terms of Use, which forbid some uses. So a checkpoint as a whole is not fully Apache. See NOTICE.