UDGAM LABS · Speech-to-text model · about 56M parameters · part of the Shruti family

Shruti Maithili

Maithili and English in one 29.7 MB file that runs on your device. Maithili comes out in Devanagari, and it keeps English words in English letters, as here: आब packet बना कऽ online बेचबाक विचार कऽ रहल छी. It runs on a phone, a laptop or in a browser, with no server and no internet connection. On a 2021 Samsung Galaxy F62 it transcribes about 9x faster than real time using 64 MB of memory.

We make no comparative claim for this model. IndicConformer Maithili is ahead on IndicVoices valid (39.3 vs 42.5) and on DiarBench's native reference (40.2 vs 51.2). With English counted in English letters, the 54-turn test shows no lead either way (45.7 vs 46.9, an interval that includes 0). What it offers is one small file that works offline. No Maithili reader has reviewed this output yet.

Shruti Maithili is part of the Shruti family, one small model per language, and is built on Cactus Whistle (Apache-2.0) via Shruti Hindi: we retrained all its weights on Maithili and English speech and added 400 Maithili vocabulary pieces.

Shruti (श्रुति) is Sanskrit for "that which is heard".

Try it: on udgamlabs.com, on your own device · demo on Hugging Face · Code and examples: github.com/udgamlabs/shruti

Lines read by AI voices from Hindi libraries, not native Maithili speakers, written by Shruti Maithili on the phone (native engine, unedited):

दरभंगामे एहि बेर मखानक दाम बढ़ि गेलै आब packet बना कऽ online बेचबाक विचार कऽ रहल छी
कोसीमे पानि बढ़ि रहल छै गामसभक लोक नाव तैयार राखू mobile charge कयलियै रातिके बाँधपर नहि जाउ

(Where they differ from the spoken lines: none; गाम सभक → गामसभक; कऽ लिअ राति कऽ बान्ह पर → कयलियै रातिके बाँधपर.)


Developer	Udgam Labs
Base model	Shruti Hindi (Udgam Labs), built on Cactus Whistle (Cactus Compute); both Apache-2.0
Parameters	About 56M (55,722,313; 8 encoder + 8 decoder layers, width 512)
File	`shruti-maithili.cact`, 29.7 MB (4-bit weights, 8-bit activations) for the Cactus needle engine
Input / output	16 kHz audio of any length → text; Maithili in Devanagari, English words in lowercase Latin when it keeps them in English
Languages	Maithili, English, and Maithili mixed with English. Other languages have their own Shruti models
Runtime	Cactus needle engine (Apache-2.0), native and WebAssembly; we measured it on a Mac, an Android phone and that phone's Chrome
Licence	Apache-2.0 for this model; see the data section for the terms of its training data

Quick start

pip install cactus-needle
hf download udgamlabs/shruti-maithili shruti-maithili.cact config.json --local-dir .

from needle import Whistle

model = Whistle("shruti-maithili.cact")
print(model.transcribe("clip.wav"))   # Maithili or English is detected automatically

The cactus-needle Python package sends anonymous usage counts by default; set NEEDLE_TELEMETRY=0 to turn that off. Command line: ./needle --model shruti-maithili.cact --audio clip.wav with the needle binary for your platform from Cactus-Compute/needle3. In the browser, the wasm/ build of the same engine runs it inside a web page; udgamlabs.com/models/shruti/maithili is a working page.

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Maithili in its on-device engine on a Mac. Measured 3 November 2026. Size is the file as published by its makers.

Model	File size	IndicVoices valid	DiarBench, English in English	DiarBench, native
Shruti Maithili	29.7 MB	42.5	45.7	51.2
IndicConformer Maithili (int8)	137.7 MB	39.3	46.9	40.2

How to read this. IndicVoices valid is IndicVoices' own held-out split: its speakers are kept out of training and its sentences blocked, but it is the same source as our training data, so it flatters us. DiarBench is real conversation from outside it: 54 turns from 15 conversations, 727 reference words, 109 of them English, in 30 turns. The "English in English" column scores against references that write English words in English (Latin script), and English words count only if written in English letters; the native column scores against references that write those English words in Devanagari.

Paired 95% intervals below are Shruti Maithili minus the other model, in word error points (1,000 resamples of clips; "by conversation" resamples whole conversations). A negative number means Shruti Maithili made fewer errors; an interval that includes 0 shows no lead either way.

IndicConformer Maithili: IndicVoices valid +1.01 to +5.73 by clip; DiarBench native +6.67 to +14.77 by clip, +6.62 to +14.84 by conversation, +6.12 to +15.56 with script ignored; DiarBench English in English -6.29 to +3.88 by clip, -5.54 to +3.23 by conversation, -2.81 to +7.26 with script ignored.
English words in English letters: 46.8% of DiarBench's 109 English words, on 54 turns. IndicConformer Maithili writes none.
One file for two languages: Shruti Maithili also transcribes English: LibriSpeech test-clean 6.04 (Shruti Hindi: 5.95).
Silence: 0 words invented on 100 silent clips.

Other Maithili speech models exist that we have not measured, and these results cover only the models listed here.

Speed and memory (offline): Samsung Galaxy F62 (2021, 6 GB RAM), native engine, 9.1x faster than real time, 64 MB peak memory (20 clips, 112.0 s of audio; slowest clip 2.7x). In that phone's Chrome, on 5 lines played into the page, it took 2.6-4.0 s on the device for 7.9-9.4 s of speech.

Where it falls short

IndicConformer Maithili is ahead or level on every test. IndicVoices valid 42.5 vs 39.3 (+1.01 to +5.73), DiarBench native 51.2 vs 40.2 (+6.67 to +14.77), DiarBench English in English 45.7 vs 46.9 (-6.29 to +3.88). Character errors on DiarBench 31.26% vs 15.02%.
A small test. 54 turns from 15 conversations, 30 of them with English, 109 English words. It can't say much either way, including whether the two are level.
Unmeasured rivals. rockerritesh/whisper-tiny-maithili (151.1 MB, about 44 MB as q8_0) and others exist. We have not measured them.
English words in Devanagari: about 34 of DiarBench's 109 by automatic alignment (shastri as शास्त्री). DiarBench also glosses some local names in Latin, so a correct native spelling can score as an error.
Browser: tested on 5 lines only, played into the page on one phone's Chrome with the language set. All 5 came out in Devanagari, and 3 kept some English letters. Not tested with a live microphone. On the other Shruti languages where we measured it, plain English loses a little accuracy with the language set.
No Maithili reader has reviewed this output yet.
Loud noise: Shruti models sometimes repeat a phrase over and over in loud noise (measured on Shruti Hindi-family models, not on Maithili).
Not tested on Maithili: noise, music, far-away or speaker-played audio, Linux, Windows and iOS. No punctuation; numbers come out as words.

How it was trained

Start: Shruti Hindi (itself a full fine-tune of Cactus Whistle). We added 400 Maithili vocabulary pieces and used Shruti Hindi's language tag for Maithili, then trained all weights with 4-bit quantisation-aware training.
Data: Maithili speech, with English words kept in English where the dataset marks them, plus English from Shruti Hindi's own training data. Test sentences were removed.

Source	Licence
AI4Bharat IndicVoices, Maithili	CC-BY-4.0
ARTPARK-IISc Vaani, Maithili (transcribed part)	CC-BY-4.0
English from Shruti Hindi's training data: LibriSpeech, Mozilla Common Voice 17 English, AI4Bharat NPTEL	CC-BY-4.0, CC0, CC-BY-4.0

Intended use and limitations

For transcribing Maithili and English on devices where privacy, cost or connectivity rule out a cloud service: voice notes, dictation, captions, field apps. If accuracy matters more than size, use a larger model. Not for high-stakes use without human review. Accuracy drops on noisy audio, strong accents, rare names, numbers and long technical English.

Licence

Apache-2.0 for this model. Shruti Maithili is a modified version of Shruti Hindi (Udgam Labs) and of Cactus Whistle by Cactus Compute (both Apache-2.0); see NOTICE. That licence covers this model, not its training data: the data section above gives each source's terms. Please credit the datasets if you build on it.

Citation

@misc{udgamlabs2026shrutimaithili,
  title  = {Shruti Maithili: offline speech-to-text for Maithili and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-maithili}
}

About Udgam Labs. Udgam Labs is an independent AI research lab. We work from first principles: before we build, we understand why things work, whether we are training a model, fine-tuning one or applying AI to a new problem. Udgam (उद्गम) is Sanskrit for origin. udgamlabs.com