Hima-ASR

Hima-ASR is a compact open speech recognition model for 11 Indian languages and English: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English.

Task	speech to text
Size	0.6B parameters (1.2 GB in bf16)
Input	16 kHz mono audio, 1 to 30 seconds per clip
Output	text in the language's own script (native), or with English words in Latin letters (mixed)
Runs on	one GPU, or CPU for short clips

Quick start

pip install "transformers>=5.13" torch torchaudio soundfile

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("Prahaladha/hima-asr", trust_remote_code=True, dtype=torch.float32)
model = model.to("cuda").eval()                       # or leave on CPU

print(model.transcribe("clip.wav", language="hi"))                 # one file -> one string
print(model.transcribe(["a.wav", "b.wav"], language="ta"))          # list -> list
print(model.transcribe("clip.wav"))                                 # the model detects the language
print(model.transcribe("clip.wav", language="hi", style="mixed"))   # English words in Latin letters

transcribe also takes 16 kHz float arrays (sampling_rate= for other rates). On a GPU it runs in bfloat16 automatically.

Argument	Values	Default
language	as bn en gu hi kn ml mr or pa ta te, or None	None (detected)
style	native (own script), mixed (English words in Latin letters)	native
batch_size	clips decoded together	16

Giving the language gives the best accuracy.

Architecture

Hima-ASR is an encoder-decoder model in three parts:

1. Audio encoder: turns 16 kHz speech into acoustic frames (12.5 per second). Initialised from the audio encoder of Qwen3-ASR-1.7B.
2. Projector: maps the audio encoder's frames into the text decoder's input space.
3. Text decoder: writes the transcript piece by piece (greedy). Initialised from Gemma 3 270M.

The text decoder can be told the language and writing style, or it can name the language itself before writing.

Benchmarks

Word error rate (WER, %, lower is better). All models are scored on the same clips with the same scorer, the language is given to each model (Sarvam through its API), and text is compared after standard normalisation (punctuation removed, Unicode NFC). Best result per row in bold.

Model	Parameters
Hima-ASR	0.6B
Indic-Transcribe-Flex	1.2B
Indic-Transcribe-Core	1.2B
Sarvam Saaras v3 (API)	not published

By language

Language	Hima-ASR	Flex	Core	Sarvam Saaras v3
English	7.5	6.4	6.9	4.0
Gujarati	10.2	10.5	10.1	10.6
Kannada	11.1	13.5	13.0	11.7
Punjabi	10.9	9.4	9.1	9.7
Bengali	12.3	10.9	10.0	10.5
Odia	12.5	12.9	12.3	11.1
Telugu	13.4	15.0	14.6	15.5
Marathi	13.9	13.4	13.4	12.0
Hindi	14.2	13.2	13.1	16.6
Malayalam	15.0	16.8	16.5	16.7
Assamese	17.0	15.8	15.2	14.7
Tamil	17.4	20.9	19.8	20.5
Average (12 languages)	12.95	13.23	12.84	12.79

Hima-ASR has the lowest error of the four in Tamil, Telugu, Malayalam and Kannada, and beats Sarvam in Hindi (14.2 vs 16.6). On the 12-language average it beats Flex and is within 0.2 points of Core (12.84) and Sarvam (12.79); Sarvam is best on English.

By test set

Test set	Language	Hima-ASR	Flex	Core	Sarvam Saaras v3
Kathbath (read speech)	Hindi	3.4	2.5	2.6	3.4
Kathbath	Bengali	8.3	5.4	5.1	7.2
Kathbath	Gujarati	7.3	5.4	5.2	6.7
Kathbath	Odia	8.3	10.0	8.7	8.6
Kathbath	Punjabi	8.1	7.1	6.9	7.7
Kathbath	Kannada	9.5	11.7	10.4	11.4
Kathbath	Marathi	11.0	8.8	8.7	9.9
Kathbath	Telugu	12.9	12.8	11.6	13.7
Kathbath	Tamil	15.4	14.1	13.8	15.0
Kathbath	Malayalam	19.9	26.8	23.6	25.2
FLEURS	Hindi	5.7	4.3	4.2	4.9
FLEURS	12 languages, average	12.7	12.7	12.3	12.0
Lahaja (many accents)	Hindi	8.8	8.1	8.0	12.1
Monsoon (conversational)	Hindi	11.1	8.0	7.6	9.9
Monsoon (conversational)	English	4.2	6.1	6.3	3.5
Svarah (Indian-accented)	English	11.4	6.8	7.8	4.7
Gram Vaani (phone-quality, rural)	Hindi	17.2	16.5	16.4	20.1

More test sets: conversational and read speech from two more public collections (SPRING-INX and MILE), 18 sets, about 1,100 to 1,500 clips each.

Test set	Language	Words	Hima-ASR	Flex	Core	Sarvam Saaras v3
SPRING-INX round 1	Assamese	21,758	35.6	38.5	37.9	38.9
SPRING-INX round 2	Assamese	25,447	33.4	35.6	34.9	37.2
SPRING-INX round 1	Bengali	25,210	26.2	26.9	26.4	26.8
SPRING-INX round 2	Bengali	23,672	27.3	27.8	27.3	28.1
SPRING-INX round 1	Gujarati	29,707	35.9	37.2	35.9	36.4
SPRING-INX round 2	Gujarati	27,419	39.6	41.1	39.9	39.9
MILE	Kannada	19,997	13.8	15.8	15.0	15.5
SPRING-INX round 1	Kannada	16,169	49.0	58.0	57.5	53.8
SPRING-INX round 1	Malayalam	17,966	47.3	49.0	47.9	47.8
SPRING-INX round 2	Malayalam	18,516	50.8	52.9	52.4	51.9
SPRING-INX round 1	Marathi	16,992	38.7	39.6	38.3	37.1
SPRING-INX round 2	Marathi	25,875	34.2	38.6	34.0	33.0
SPRING-INX round 1	Odia	21,014	35.3	40.0	38.8	37.3
SPRING-INX round 1	Punjabi	24,950	30.3	32.8	30.9	30.6
SPRING-INX round 2	Punjabi	27,150	33.1	34.7	33.4	33.3
MILE	Tamil	10,966	14.0	18.8	17.3	18.7
SPRING-INX round 1	Tamil	25,941	36.4	46.5	44.7	39.6
SPRING-INX round 2	Tamil	19,107	37.5	50.5	48.1	40.9

Hima-ASR has the lowest error of the four on 15 of these 18 sets; it is ahead of Core on 15, of Sarvam on 16 and of Flex on all 18.

Files

File	What it is
model.safetensors	weights (bf16)
config.json, configuration_hima_asr.py	configuration
modeling_hima_asr.py	model code (loaded with trust_remote_code=True)
preprocessor_config.json	audio feature settings
tokenizer.json, tokenizer_config.json	text decoder tokenizer

Limitations

• Clips longer than 30 seconds should be split before transcription.
• Accuracy is lower on phone-quality and noisy audio than on clean speech.
• Without a language hint, the model can occasionally write a clip in the script of a related language.
• Proper nouns and rare words are the most common errors.

License

• Hima-ASR's own parts (projector, training changes, inference code): Apache 2.0, see LICENSE.
• Audio encoder: initialised from Qwen3-ASR-1.7B, Apache 2.0.
• Text decoder: initialised from Gemma 3 270M. Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms, including its use restrictions. Hima-ASR is a Model Derivative of Gemma, so any use or redistribution of Hima-ASR must follow those terms. See NOTICE.