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

Shruti Konkani

Konkani and English in one 29.7 MB file. On Konkani conversations, about 15% fewer errors than IndicConformer Konkani, a model almost five times its size. Konkani comes out in Devanagari, and about half of the English words stay in English letters: बांगडे आनी सुंगटां fridge भितर दवरल्यात. 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 12x faster than real time using 64 MB of memory.

The scope matters. On DiarBench (held out, all 90 usable turns from 23 conversations, 71 of them with English words), counting English words only in English letters, its word error rate is 40.0 against IndicConformer Konkani's 47.0, and that holds by conversation and with script ignored. Against the native reference on the same turns IndicConformer Konkani is well ahead (37.3 vs 47.3), and on IndicVoices valid too (33.4 vs 41.4). It is the one rival we measured. It writes Devanagari Konkani, not Romi or Kannada script. No Konkani reader has reviewed this output yet.

Shruti Konkani 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 Konkani and English speech and added 400 Konkani 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 Marathi libraries, not native Konkani speakers, written by Shruti Konkani on the phone (native engine, unedited):

आयज घरा नुस्त्याचें हुमण आनी शीत जेवपाक आसा बांगडे आनी सुंगटां fridge भितर दवरल्यात
season सुरू जाला आयज शॅकच्या साहित्य कळमुटाक व्हरता फुडल्या सुमानान tourist येवपाक लागतले

(Where they differ from the spoken lines: none; shack चें → शॅकच्या; कळंगुटाक व्हरतां → कळमुटाक व्हरता; सुमानांत → सुमानान.)


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

from needle import Whistle

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

Results

Word error rate (%, lower is better), the same clips and scorer for every model; Shruti Konkani 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 Konkani	29.7 MB	41.4	40.0	47.3
IndicConformer Konkani (int8)	137.7 MB	33.4	47.0	37.3

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: 90 turns from 23 conversations, 1,457 reference words, 278 of them English, in 71 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 Konkani minus the other model, in word error points (1,000 resamples of clips; "by conversation" resamples whole conversations). A negative number means Shruti Konkani made fewer errors; an interval that includes 0 shows no lead either way.

IndicConformer Konkani: IndicVoices valid +6.10 to +10.00 by clip; DiarBench native +6.41 to +13.71 by clip, +6.60 to +13.63 by conversation, +6.69 to +12.53 with script ignored; DiarBench English in English -10.64 to -3.57 by clip, -10.86 to -3.50 by conversation, -6.96 to -0.90 with script ignored.
English words in English letters: 49.6% of DiarBench's 278 English words, on 90 turns. IndicConformer Konkani writes none.
One file for two languages: Shruti Konkani also transcribes English: LibriSpeech test-clean 6.0 (Shruti Hindi: 5.95).
Silence: 0 words invented on 100 silent clips.

Other Konkani 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, 11.7x faster than real time, 64 MB peak memory (20 clips, 187.5 s of audio; slowest clip 4.9x). In that phone's Chrome, on 4 lines played into the page, it took 3.1-3.9 s on the device for 8.3-8.9 s of speech.

Where it falls short

Devanagari only. Goan Konkani is widely written in Romi too, and this model does not write it. Reubencf/konkani-whisper-stt targets Romi; we have not measured it.
The native reference shows the gap. IndicConformer Konkani is ahead on DiarBench native (37.3 vs 47.3; +6.41 to +13.71) and on IndicVoices valid (33.4 vs 41.4). Character errors: DiarBench 25.38% vs 11.58%, IndicVoices valid 22.77% vs 10.1%.
One rival measured. No independent read test was available either: IndicVoices valid shares a source with our training data.
English words in Devanagari: about 77 of DiarBench's 278 by automatic alignment (so as सो). 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 Konkani 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 Konkani).
Not tested on Konkani: 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 Konkani vocabulary pieces and used Shruti Hindi's language tag for Konkani, then trained all weights with 4-bit quantisation-aware training.
Data: Konkani 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, Konkani	CC-BY-4.0
ARTPARK-IISc Vaani, Konkani (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 Konkani 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 Konkani 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{udgamlabs2026shrutikonkani,
  title  = {Shruti Konkani: offline speech-to-text for Konkani and English in 29.7 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/shruti-konkani}
}

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