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

Shruti Manipuri

Manipuri in Meetei Mayek and English in one 29.7 MB file that runs on your device. Manipuri comes out in Meetei Mayek, and about 60% of DiarBench's 243 English words stay in English letters: ꯍꯦꯡ ꯑꯌꯨꯛ ꯈꯨꯃꯟ ꯃꯄꯥꯛꯇ boxing practice ꯂꯩꯒꯅꯤ. 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.

A pointer, not a result. On 44 conversation turns from 20 conversations (DiarBench, held out, 39 turns with English words), counting English words only in English letters, it has fewer errors than IndicConformer Manipuri (48.2 vs 65.4), but that is under the size we'd trust, and the interval is wide. Against the native reference on the same turns IndicConformer Manipuri is far ahead (53.5 vs 66.0), and on IndicVoices valid too (31.2 vs 36.0). It writes Meetei Mayek, not Bengali script. No Manipuri reader has reviewed this output yet.

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

Shruti (ꯁ꯭ꯔꯨꯇꯤ) 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 Manipuri speaker has checked it:

ꯍꯦꯡ ꯑꯌꯨꯛ ꯈꯨꯃꯟ ꯃꯄꯥꯛꯇ boxing practice ꯂꯩꯒꯅꯤ


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

from needle import Whistle

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

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Manipuri 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 Manipuri	29.7 MB	36.0	48.2	66.0
IndicConformer Manipuri (int8)	137.7 MB	31.2	65.4	53.5

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: 44 turns from 20 conversations, 815 reference words, 243 of them English, in 39 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 Meetei Mayek.

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

IndicConformer Manipuri: IndicVoices valid +2.07 to +7.70 by clip; DiarBench native +7.02 to +18.67 by clip, +6.39 to +20.31 by conversation; DiarBench English in English -26.78 to -7.98 by clip, -31.04 to -8.05 by conversation.
English words in English letters: 60.9% of DiarBench's 243 English words, on 44 turns. IndicConformer Manipuri writes none.
One file for two languages: Shruti Manipuri also transcribes English: LibriSpeech test-clean 5.97 (Shruti Hindi: 5.95).
Silence: 0 words invented on 100 silent clips.

Other Manipuri 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, 11.2x faster than real time, 64 MB peak memory (20 clips, 135.9 s of audio; slowest clip 3.9x). In that phone's Chrome, on 4 lines played into the page, it took 2.0-2.9 s on the device for 7.3-8.0 s of speech.

Where it falls short

A small test. 44 turns from 20 conversations; the interval is wide (-26.78 to -7.98), and no script-ignored check exists for Meetei Mayek.
The native reference shows the gap. DiarBench native 66.0 vs 53.5. Character errors there 40.29% vs 17.48%, the highest in this group. IndicVoices valid 36.0 vs 31.2.
Meetei Mayek only. It does not write Manipuri in Bengali script.
Sometimes the whole line comes out in English letters: 4 of 200 IndicVoices clips and 2 of 44 DiarBench turns, with the engine's own language choice.
English words in Meetei Mayek: about 33 of DiarBench's 243 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 test clips only, played into the page on one phone's Chrome with the language set. All 4 came out in Meetei Mayek. 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 Manipuri 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 Manipuri).
Not tested on Manipuri: 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 Manipuri vocabulary pieces and used Shruti Hindi's language tag for Manipuri, then trained all weights with 4-bit quantisation-aware training.
Data: Manipuri 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, Manipuri	CC-BY-4.0
ARTPARK-IISc Vaani, Manipuri (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 Manipuri 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 Manipuri 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{udgamlabs2026shrutimanipuri,
  title  = {Shruti Manipuri: offline speech-to-text for Manipuri and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-manipuri}
}

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