udgamlabs/shruti-dogri

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

Shruti Dogri

Dogri and English in one 29.7 MB file that runs on your device. Dogri comes out in Devanagari, and it keeps English words in English letters, as here: कटड़े आह्‌ली बस हून चलने आह्‌ली ऐ ticket लेई लैओ. 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 10x faster than real time using 64 MB of memory.

We make no comparative claim for this model. IndicConformer Dogri is ahead on IndicVoices valid, the one sizeable test (38.5 vs 43.5), and the Dogri conversation test is only 24 turns, too few to say anything either way. What it offers is one small file that works offline. No Dogri reader has reviewed this output yet.

Shruti Dogri 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 Dogri and English speech and added 400 Dogri 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 Dogri speakers, written by Shruti Dogri on the phone (native engine, unedited):

कटड़े आह्‌ली बस हून चलने आह्‌ली ऐ ticket लेई लैओ समान पिच्छें रक्खो बैठने दी seat अजे ऐ
मेरा जागत फौज दी परतियै आस्तै जम्मू गेदा ऐ अज्ज ओह्‌दा physical test ऐ दवा करेओ पास होई जा

(Where they differ from the spoken lines: आह्ली bus हुन → आह्‌ली बस हून; आह्ली → आह्‌ली; सामान → समान; अजें है → अजे ऐ; भर्ती → परतियै; ओह्दा → ओह्‌दा; दुआ → दवा.)

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-dogri.cact, 29.7 MB (4-bit weights, 8-bit activations) for the Cactus needle engine
Input / output	16 kHz audio of any length → text; Dogri in Devanagari, English words in lowercase Latin when it keeps them in English
Languages	Dogri, English, and Dogri 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-dogri shruti-dogri.cact config.json --local-dir .

from needle import Whistle

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

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Dogri 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 Dogri	29.7 MB	43.5	34.5	37.0
IndicConformer Dogri (int8)	137.7 MB	38.5	39.7	37.4

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: 24 turns from 17 conversations, 479 reference words, 40 of them English, in 16 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 Dogri minus the other model, in word error points (1,000 resamples of clips; "by conversation" resamples whole conversations). A negative number means Shruti Dogri made fewer errors; an interval that includes 0 shows no lead either way.

• IndicConformer Dogri: IndicVoices valid +2.58 to +7.27 by clip; DiarBench native -5.31 to +5.32 by clip, -6.15 to +5.64 by conversation, -1.74 to +10.86 with script ignored; DiarBench English in English -10.99 to +1.78 by clip, -12.13 to +1.99 by conversation, -7.30 to +6.09 with script ignored.
• One file for two languages: Shruti Dogri also transcribes English: LibriSpeech test-clean 6.1 (Shruti Hindi: 5.95).
• Silence: 0 words invented on 100 silent clips.

Other Dogri 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, 9.5x faster than real time, 64 MB peak memory (20 clips, 139.1 s of audio; slowest clip 4.7x). In that phone's Chrome, on 4 lines played into the page, it took 1.5-1.6 s on the device for 8.8-9.2 s of speech.

Where it falls short

• The test can't support a claim. DiarBench Dogri is 24 turns from 17 conversations, 479 words, 40 English words. No independent read set was available.
• IndicConformer Dogri is ahead on IndicVoices valid (38.5 vs 43.5; +2.58 to +7.27). Character errors: 26.59% vs 17.91% there, 25.64% vs 16.98% on DiarBench.
• One rival measured. Whisper does not list Dogri. Other Dogri models exist that we have not measured.
• English words in Devanagari: about 12 of DiarBench's 40 by automatic alignment. DiarBench also glosses some local names in Latin, so a correct native spelling can score as an error.
• Browser: tested on 4 lines only, played into the page on one phone's Chrome with the language set. All 4 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 Dogri 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 Dogri).
• Not tested on Dogri: 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 Dogri vocabulary pieces and used Shruti Hindi's language tag for Dogri, then trained all weights with 4-bit quantisation-aware training.
• Data: Dogri 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, Dogri	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 Dogri 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 Dogri 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{udgamlabs2026shrutidogri,
  title  = {Shruti Dogri: offline speech-to-text for Dogri and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-dogri}
}

---

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