udgamlabs/shruti-santali

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

Shruti Santali

Santali in Ol Chiki and English in one 29.7 MB file that runs on your device. Santali comes out in Ol Chiki, and some English words stay in English letters, as here: ᱤᱱᱟᱹ ᱵᱚᱭᱦᱟ ᱡᱟᱢᱥᱮᱫᱽᱯᱩᱨ ᱨᱮ ᱢᱤᱫᱴᱟᱝ company ᱨᱮ ᱠᱟᱹᱢᱤᱭᱮᱫᱟᱭ. 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 13x faster than real time using 64 MB of memory.

On our two tests we found no clear difference from IndicConformer Santali, a model almost five times its size. On DiarBench (99 turns, native reference) 33.6 vs 34.2, on IndicVoices valid 40.3 vs 42.6, and both intervals include 0. IndicConformer Santali has fewer character errors on both (11.7% vs 15.88%, 13.3% vs 16.88%). Don't read this as a lead: the size is the difference. Its English test is thin, 63 English words. No Santali reader has reviewed this output yet.

Shruti Santali 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 Santali and English speech and added 700 Santali vocabulary pieces.

Shruti (ᱥᱨᱩᱛᱤ) 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

An example of how it writes, from a test clip read by an AI voice, not real speech. It is the phone's unedited output, and no Santali speaker has checked it:

ᱤᱱᱟᱹ ᱵᱚᱭᱦᱟ ᱡᱟᱢᱥᱮᱫᱽᱯᱩᱨ ᱨᱮ ᱢᱤᱫᱴᱟᱝ company ᱨᱮ ᱠᱟᱹᱢᱤᱭᱮᱫᱟᱭ

Developer	Udgam Labs
Base model	Shruti Hindi (Udgam Labs), built on Cactus Whistle (Cactus Compute); both Apache-2.0
Parameters	About 56M (55,875,913; 8 encoder + 8 decoder layers, width 512)
File	shruti-santali.cact, 29.7 MB (4-bit weights, 8-bit activations) for the Cactus needle engine
Input / output	16 kHz audio of any length → text; Santali in Ol Chiki, English words in lowercase Latin when it keeps them in English
Languages	Santali, English, and Santali 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-santali shruti-santali.cact config.json --local-dir .

from needle import Whistle

model = Whistle("shruti-santali.cact")
print(model.transcribe("clip.wav"))   # Santali 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-santali.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/santali is a working page.

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Santali 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 Santali	29.7 MB	40.3	33.0	33.6
IndicConformer Santali (int8)	137.7 MB	42.6	35.4	34.2

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: 99 turns from 28 conversations, 2,897 reference words, 63 of them English, in 34 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 Ol Chiki.

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

• IndicConformer Santali: IndicVoices valid -5.29 to +0.52 by clip; DiarBench native -3.16 to +2.03 by clip, -2.67 to +1.61 by conversation; DiarBench English in English -4.94 to +0.25 by clip, -4.60 to +0.10 by conversation.
• English words in English letters: 36.5% of DiarBench's 63 English words, on 99 turns. IndicConformer Santali writes none.
• One file for two languages: Shruti Santali also transcribes English: LibriSpeech test-clean 5.98 (Shruti Hindi: 5.95).
• Silence: 0 words invented on 100 silent clips.

Other Santali 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, 12.6x faster than real time, 64 MB peak memory (20 clips, 202.2 s of audio; slowest clip 3.5x). In that phone's Chrome, on 4 lines played into the page, it took 2.4-3.9 s on the device for 6.4-9.4 s of speech.

Where it falls short

• No lead over the one rival measured. DiarBench native -3.16 to +2.03, DiarBench English in English -4.94 to +0.25, IndicVoices valid -5.29 to +0.52. Character errors are higher: 15.88% vs 11.7% and 16.88% vs 13.3%.
• Ol Chiki only. It does not write Santali in Devanagari, Bengali, Odia or Latin script, and no script-ignored check exists for Ol Chiki.
• A thin English test. 63 English words in 2,897, of which 36.5% come out in English letters.
• English words in Ol Chiki: about 30 of DiarBench's 63 by automatic alignment (doctor as ᱰᱟᱠᱛᱟᱨ). DiarBench also glosses some local names in Latin, so a correct native spelling can score as an error.
• Browser: tested on 4 test clips only, played into the page on one phone's Chrome with the language set. All 4 came out in Ol Chiki. These are test clips read by an AI voice, not real speech. 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 Santali 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 Santali).
• Not tested on Santali: 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 700 Santali vocabulary pieces and used Shruti Hindi's language tag for Santali, then trained all weights with 4-bit quantisation-aware training.
• Data: Santali 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, Santali	CC-BY-4.0
ARTPARK-IISc Vaani, Santali (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 Santali 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 Santali 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{udgamlabs2026shrutisantali,
  title  = {Shruti Santali: offline speech-to-text for Santali and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-santali}
}

---

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