UDGAM LABS · Speech-to-text model · 55M parameters

Shruti

Shruti turns Hindi, English and Hinglish speech into text, offline, from one 29.5 MB file. Hindi comes out in Devanagari and English words stay in English, the way people actually type Hinglish: मेरा order कब आएगा. It runs on a phone, a laptop or inside a web page, with no server and no internet connection. On a 2021 Samsung Galaxy F62 it transcribes a sentence in about a second using 80 MB of memory.

Shruti is built on Cactus Whistle (Apache-2.0): we retrained all 55M weights on 2,732 hours of openly licensed Indian speech and added 714 Devanagari vocabulary pieces.

Shruti (श्रुति) is Sanskrit for "that which is heard".

Try it: in your browser, on your own device · demo on Hugging Face · Code and examples: github.com/udgamlabs/shruti · Technical report: udgamlabs.com/research/shruti

Audio (Hinglish)                       Shruti writes
"नहीं but वो safely बच रहे..."     ->  नहीं but वो safely बच रहे हों तो क्या अच्छी है
"improve होता है"                  ->  improve होता है
English: "Nipah virus belongs to the"  nipa virus belongs to the

Developer	Udgam Labs
Base model	Cactus Whistle (Cactus Compute), Apache-2.0
Parameters	55M (8 encoder + 8 decoder layers, width 512)
Files	shruti.cact, 29.5 MB (4-bit weights, 8-bit activations) for the Cactus needle engine; shruti.safetensors, 222 MB (full-precision fp32 weights, config in the file's metadata)
Input / output	16 kHz audio of any length → text; Hindi in Devanagari, English words in lowercase Latin
Languages	Hindi, English, Hinglish (code-mixed). Whistle's other languages (German, French, Spanish, Italian, Dutch, Polish) were not evaluated and are probably weaker after our training
Runtime	Cactus needle engine (Apache-2.0): macOS, Linux, Windows, Android, iOS, WebAssembly
Licence	Apache-2.0

Quick start

Python (macOS, Linux, Windows):

pip install cactus-needle
hf download udgamlabs/Shruti shruti.cact --local-dir .

from needle import Whistle

model = Whistle("shruti.cact")
print(model.transcribe("clip.wav"))                 # the language is detected automatically
# {'text': 'नहीं but वो safely बच रहे हों तो क्या अच्छी है', 'language': 'hi', ...}
print(model.transcribe("english.wav", language="en"))  # only for speech that is all English (see below)

transcribe also takes keywords=[...] (names or terms to favour) and word_timestamps=True. The cactus-needle Python package sends anonymous usage counts (function name, versions, OS) by default; set NEEDLE_TELEMETRY=0 to turn that off. The command-line engine and the WebAssembly build make no network calls.

Command line: download the needle binary for your platform from Cactus-Compute/needle3 (folders macos-arm64, linux-x86_64, windows-x86_64, android-arm64, …), then:

./needle --model shruti.cact --audio clip.wav                     # language detected automatically
./needle --model shruti.cact --audio talk.wav --audio-stream      # live, word by word
./needle --model shruti.cact --audio english.wav --audio-language en   # all-English speech

In the browser: the wasm/ build of the same engine runs Shruti inside a web page (single-threaded, no special server headers). See udgamlabs.com/models/shruti for a working page.

In an app (Android, iOS): link the engine's C library (needle.h: needle_load, needle_transcribe, needle_stream_transcribe_process) and ship shruti.cact with the app. Use 2 threads on phones with two fast cores.

Results

Word error rate (%, lower is better). Every model was run by us on the same test sets with the same text normaliser and scorer (Monsoon uses the official OIWER lattice scorer of the Hugging Face Open ASR leaderboard). Shruti was run in the on-device engine (5-beam, language set); the other models on a GPU or a Mac with the decoding settings listed in eval/comparison.md, with exact versions. Measured 2026-10-18/19. n/c: not comparable (the model writes Roman script only); –: not measured.

Model	File size	Monsoon Hindi	FLEURS Hindi	Lahaja (Hindi accents)	MUCS Hinglish	Svarah (Indian English)	LibriSpeech (English)	Words invented on 100 silent clips
Shruti	29.5 MB	20.46	19.92	26.28	35.21	22.47	5.95	0
Koyal-hi-120M	481 MB	11.64	9.56	18.70	57.51	Hindi only	Hindi only	0
IndicConformer Hindi (fp32)	482 MB	15.31	11.31	13.18	64.48	Hindi only	Hindi only	88
Vosk small Hindi + Indian English	139 MB	35.75	18.07	26.60	61.14	38.51	30.18	1
Whisper-Hindi2Hinglish-Swift	296 MB	57.16 (Roman script)	n/c	n/c	n/c	20.16	12.60	1,676
Whisper-small	971 MB	128.25	59.42	117.5	–	–	–	–
whisper-base-hinglish (fp32)	290 MB	65.66	–	–	87.05 (Roman script); 43.45 romanised	27.11 ¹	15.53 ¹	500
whisper-tiny-hinglish-v2 (fp32)	151 MB	70.60	–	–	87.73 (Roman script); 44.94 romanised	34.50 ¹	23.92 ¹	500
Whisper large-v3 (fp16)	3,087 MB	33.29	28.88	31.19 ²	79.94 ²	7.39 ²	2.25 ²	538

¹ First 1,500 Svarah and first 1,000 LibriSpeech clips; Shruti on the same clips: 21.35 and 6.36. These two models write everything in Roman script, so on MUCS we also scored both sides after converting them to Roman script with the same function ("romanised"); Shruti scores 34.32 that way. Their model card reports 28.4 and 32.0 on a 250-clip sample with the author's own romaniser, which we did not run. Both write the same short sign-off on every silent clip.

² Whisper large-v3 (1.55B parameters; fp16 on a Mac GPU, beam 5, language set) on the first 300 Lahaja, Svarah and LibriSpeech clips and the first 600 MUCS clips; Shruti on the same clips: 22.87, 21.35, 5.76 and 45.3. On Monsoon and FLEURS both ran on every clip; in that run Shruti scored 20.46 and 20.14 (FLEURS with the newest engine build; 19.92 in the table is the build of 18 October).

• Hinglish: Shruti has the lowest Hinglish word error rate of the open models we measured (MUCS 2021, same harness), and it stays ahead when both sides are converted to Roman script (34.32 vs 43.45 and 44.94 for whisper-base-hinglish and whisper-tiny-hinglish-v2; with the most forgiving converter the lead is narrow, 31.37 vs 32.66 and 34.22).
• Whisper large-v3: Shruti is 100× smaller than OpenAI's Whisper large-v3 (29.5 MB vs 3,087 MB), and more accurate on Hindi on the same clips: Monsoon 20.46 vs 33.29, FLEURS 20.14 vs 28.88, Lahaja 22.87 vs 31.19 (first 300 clips). Whisper is far better on English: Svarah 7.39 vs 21.35 and LibriSpeech 2.25 vs 5.76 (first 300 clips each). Whisper ran with plain beam search, without its temperature fallback, which causes some repetition loops; with those clips removed Shruti still leads on Hindi (FLEURS 19.9 vs 26.4, Lahaja 22.9 vs 28.7). The leaderboard's own Whisper large-v3 Monsoon number (28.17, its decoding) is also behind Shruti's 20.46.
• Monsoon is the public split of the Hindi test set used by the Hugging Face Open ASR leaderboard. Published leaderboard numbers for context: Whisper large-v3 28.17 (1.55B parameters), Voxtral Mini 3B 22.9, Qwen3-ASR-0.6B 18.13, IndicConformer-600M 8.47. Shruti has not been submitted to the leaderboard, so it has no official rank, and the full test set may score a few points differently from the public split.
• MUCS (Hindi-English code-mixed lectures): references mostly write English words in Latin script, as Shruti does. Models that write English in Devanagari, or everything in Roman script, are penalised on it; that is the Hinglish skill.

Speed and memory (offline):

Device	Result
Samsung Galaxy F62 (Exynos 9825, 6 GB RAM, Android 13), 2 threads	a ~7 s sentence in ~1 s including model load; 60 s of audio in 9.5 s (6.3× real time); peak memory 80 MB
Web browser engine (WebAssembly, one thread, Node on Apple M-series)	a 9.6 s clip in 0.45 s; transcripts identical to the native engine
Hugging Face Space (shared 2-core CPU)	a 2.8 s clip in 0.35 s

On the phone, 42 of 51 test transcripts were identical to the Mac's; the rest differed by a letter or two.

Language setting. Leave the language unset: automatic detection scores the same as setting it on Hindi, Hinglish, accented Hindi and US English. Set en only for speech that is entirely English with an Indian accent, which detection often treats as Hindi (Svarah subset: 21.1 with en vs 25.8 detected). The results table above sets the language (hi for the Hindi and Hinglish sets, en for the English sets).

Where it falls short

• Pure Hindi accuracy. Hindi-only models about 16× larger are far more accurate: Koyal-hi-120M makes roughly half as many errors on Monsoon and FLEURS, and IndicConformer half as many on accented Hindi (Lahaja). Choose them if you only need Hindi and have the memory.
• English. OpenAI's Whisper large-v3 (100× larger) is far more accurate on English: on the same clips, Svarah 7.39 vs 21.35 and LibriSpeech 2.25 vs 5.76. On Indian-accented English Shruti (Svarah 22.5) is also behind Whisper-Hindi2Hinglish-Swift (20.2), which is 10× larger and writes Hindi in Roman script only.
• Spelling and grammar look-alikes: है/हैं, मैं/में, की/के, and formal versus spoken forms (पे → पर, वो → वह) are common errors.
• Names it has never heard are spelt by sound: in our own phone test it wrote its name, Shruti, as "सूती" and "shooty", and Udgam Labs as "outgum labs". Rare names and technical terms are often misheard; pass them as keywords to help.
• Noise and music were not tested; only silent clips (0 invented words on 100 clips of digital silence and very faint hiss). Long audio is processed in 30 s windows.
• Numbers are written as words, the way they are spoken. There is no punctuation or capitalisation.
• Platforms: we measured Shruti on a Mac (Apple silicon), Linux (x86), an Android phone and WebAssembly. The engine also runs on Windows and iOS, but we have not tested Shruti there.

How it was trained

• Start: Cactus Whistle (55M, Apache-2.0), trained by Cactus Compute on European languages. We added 714 Devanagari vocabulary pieces and a Hindi language tag, then trained all weights (not an adapter) with 4-bit quantisation-aware training, so the exported 29.5 MB file behaves like the trained model.
• Data: 2,732 hours of speech, every source licensed for commercial use. English words in Hindi transcripts are written in Latin script using each dataset's own annotations (for example AI4Bharat's [english] tags in IndicVoices). Test sentences were removed by text and audio hash.

Source	Licence
AI4Bharat IndicVoices, Hindi	CC-BY-4.0
ARTPARK-IISc Vaani, Hindi (transcribed train split)	CC-BY-4.0
AI4Bharat Shrutilipi, Hindi	CC-BY-4.0
AI4Bharat Kathbath, Hindi	CC-BY-4.0
AI4Bharat Rural Women ASR v2, Hindi	CC-BY-4.0
Google FLEURS, Hindi train split	CC-BY-4.0
MUCS 2021 Hindi-English (via dianavdavidson/MUCS-Hinglish)	CC-BY-4.0 on the card; the original corpus is CC-BY-SA-4.0
AI4Bharat NPTEL (Indian-English lectures)	CC-BY-4.0
Mozilla Common Voice 17, English with Indian / South Asian accents	CC0
LibriSpeech	CC-BY-4.0

• Recipe: one spelling per word for training targets; trained from Whistle's weights, then continued with more Indian English.

The technical report describes the vocabulary extension, the training method, the data and the evaluation in full, with the source of every number.

Intended use and limitations

For transcribing Hindi, English and Hinglish speech on devices where privacy, cost or connectivity rule out a cloud service: voice notes, dictation, voice commands, captions, field apps. Not for high-stakes use (medical, legal) without human review. Accuracy drops on noisy audio, strong accents, rare names and long technical English.

Licence

Apache-2.0. Shruti is a modified version of Cactus Whistle by Cactus Compute (Apache-2.0); see NOTICE. The training data licences are listed above; please credit the datasets if you build on this work.

Citation

@misc{udgamlabs2026shruti,
  title  = {Shruti: offline speech-to-text for Hindi, English and Hinglish in 29.5 MB},
  author = {{Udgam Labs}},
  year   = {2026},
  url    = {https://huggingface.co/udgamlabs/Shruti}
}

@techreport{udgamlabs2026shrutireport,
  title       = {Shruti: offline speech-to-text for Hindi, English and Hinglish in 29.5 MB},
  author      = {{Udgam Labs}},
  institution = {Udgam Labs},
  type        = {Technical report},
  year        = {2026},
  url         = {https://udgamlabs.com/research/shruti}
}

---

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