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

Shruti Sindhi

Sindhi, written in Devanagari, and English in one 29.7 MB file that runs on your device. Sindhi comes out in Devanagari, and it keeps English words in English letters, as here: छोकिरे जो biodata ऐं photo मोकिलियो. 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 11x faster than real time using 64 MB of memory.

It writes Sindhi in Devanagari, not Perso-Arabic script, so readers of Perso-Arabic Sindhi can't use its output. We make no comparative claim for this model. IndicVoices valid is level with IndicConformer Sindhi (33.0 vs 33.1), on a set that shares a source with our training data, and the conversation test is only 27 turns. What it offers is one small file that works offline. No Sindhi reader has reviewed this output yet.

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

छोकिरे जो biodata ऐं photo मोकिलियो अठमी पोसो न सिब phone कजो पोइ मिलण जो programme ठाहींदासीं
सुभाणे थधिड़ी आहे चूल्ही न ॿाहिरींदासीं अॻु राति जो लोला ऐं दही fridge में रखो

(Where they differ from the spoken lines: अथमि पसंद अचे त → अठमी पोसो न सिब; चुल्हि → चूल्ही; ॿारींदासीं अॼु → ॿाहिरींदासीं अॻु.)


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

from needle import Whistle

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

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Sindhi 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 Sindhi	29.7 MB	33.0	34.2	38.0
IndicConformer Sindhi (int8)	137.7 MB	33.1	38.7	35.8

How to read this. Whisper is not in this table: it hardly writes Devanagari on these sets, so its error rates would measure script, not accuracy. 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: 27 turns from 17 conversations, 971 reference words, 51 of them English, in 12 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 Sindhi minus the other model, in word error points (1,000 resamples of clips; "by conversation" resamples whole conversations). A negative number means Shruti Sindhi made fewer errors; an interval that includes 0 shows no lead either way.

IndicConformer Sindhi: IndicVoices valid -2.45 to +2.39 by clip; DiarBench native -1.53 to +6.04 by clip, -1.43 to +4.91 by conversation, +1.99 to +8.99 with script ignored; DiarBench English in English -8.37 to +0.10 by clip, -8.46 to -0.36 by conversation, -4.55 to +4.21 with script ignored.
One file for two languages: Shruti Sindhi also transcribes English: LibriSpeech test-clean 6.05 (Shruti Hindi: 5.95).
Silence: 0 words invented on 100 silent clips.

Other Sindhi 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, 10.8x faster than real time, 64 MB peak memory (20 clips, 212.5 s of audio; slowest clip 4.5x). In that phone's Chrome, on 4 lines played into the page, it took 2.7-3.8 s on the device for 7.1-9.3 s of speech.

Where it falls short

Devanagari only. Readers of Perso-Arabic Sindhi can't use the output. FLEURS Sindhi is in Arabic script and was left out for this reason, so there is no independent read test.
A small test. DiarBench Sindhi is 27 turns from 17 conversations with 51 English words, too few to say much.
The level result is on the flattering set. IndicVoices valid (-2.45 to +2.39) shares a source with our training data. Character errors there 19.9% vs 12.41%, and on DiarBench 21.64% vs 14.33%.
English words in Devanagari: about 7 of DiarBench's 51 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 2 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 Sindhi 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 Sindhi).
Not tested on Sindhi: 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 Sindhi vocabulary pieces and used Shruti Hindi's language tag for Sindhi, then trained all weights with 4-bit quantisation-aware training.
Data: Sindhi 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, Sindhi	CC-BY-4.0
ARTPARK-IISc Vaani, Sindhi (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 Sindhi 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 Sindhi 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{udgamlabs2026shrutisindhi,
  title  = {Shruti Sindhi: offline speech-to-text for Sindhi and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-sindhi}
}

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