Shruti
Speech to text that runs on your device, with no server: one small model per language, each with English.
Speech to text in one small file per language: Shruti Hindi (Hindi, English and Hinglish, 29.5 MB) and Shruti Marathi (Marathi and English, 29.7 MB). Each writes its language in Devanagari and English words in English, the way people type, and runs offline: on a phone, on a laptop or in a browser tab.
Try it now · in your browser · free, no login
Speak Hindi, English or both. Shruti writes it down on your device.
Shruti · checking this device
Audio never leaves your device. Shruti and its open-source engine run inside this page.
(or drop one on the transcript)
Load Shruti, then press Record and speak Hindi, English or both.
The language is detected for you. English words stay in English (Latin script) and are underlined here; Hindi is written in Devanagari. While you speak, words appear once they have settled, a second or two behind you; when you stop, Shruti reads the whole recording once more for the final text.
Shruti Marathi · 29.7 MB · Marathi and English
The most accurate open Marathi speech-to-text model under 100 MB that we measured.
Shruti Marathi writes Marathi in Devanagari and keeps English words in English, the way people type: उद्या सकाळी नऊ वाजता physics ची exam आहे. Every other open model under 100 MB we measured gets more Marathi words wrong than right. It is within 1.33× of IndicConformer Marathi’s error rate on FLEURS (29.9 against 22.4) at about a fifth of its size, and on conversations that mix in English it makes fewer errors than that model when English words are written in English. It runs offline on a five-year-old phone, 12× faster than real time in 64 MB of memory with the native engine, and it invented no words on 100 silent clips.
| Model | File | FLEURS | Kathbath | DiarBench | IndicVoices | SPRING-INX | Vaani |
|---|---|---|---|---|---|---|---|
| Shruti Marathi | 29.7 MB | 29.9 | 24.5 | 38.2 | 22.2 | 42.9 | 32.2 |
| Whisper tiny | 44 MB | 273.1 | 278.1 | 373.9 | 321.8 | 375.6 | 338.7 |
| Whisper base | 82 MB | 177.9 | 193.2 | 402.9 | 450.2 | 420.6 | 412.3 |
| marathi-conformer-33m-moe | 136.7 MB | 41.0 | 32.2 | 76.7 | 66.0 | 77.3 | 64.0 |
| IndicConformer Marathi | 137.7 MB | 22.4 | 14.1 | 41.5 | 27.6 | 53.0 | 44.0 |
Word error rate %, lower is better; 200 clips per set, the same scorer for every model, measured 30 October to 2 November 2026. FLEURS and Kathbath are read Marathi; the other four are conversations that mix in English, scored with English words written in English. With the English words dropped from those references, IndicConformer Marathi is ahead (DiarBench 32.3 against 40.9). Whisper tiny and base add words, so their rates pass 100%.
Where it falls short
- On pure Marathi, IndicConformer Marathi (about 5× larger) makes about a quarter fewer errors on FLEURS and about 40% fewer on Kathbath.
- Spelling variants (उशिरा/उशीरा) and joined words count as errors and are common; some English loanwords come out in Devanagari.
- It does not do Hindi: pick Hindi above for Shruti Hindi.
- Noise, music and speaker-played audio were not tested.
Shruti Marathi on Hugging Face ↗ · the model file, the full results and how each model was run
Shruti Hindi · the result
100× smaller than OpenAI’s Whisper large-v3, and better at Hindi.
On the same Hindi test clips, Shruti makes fewer word errors than Whisper large-v3, a 3,087 MB model: Monsoon 20.5 against 33.3, FLEURS 20.1 against 28.9, Lahaja 22.9 against 31.2. On English, Whisper is clearly better (Indian English 21.4 against 7.4 on the same clips). It is also the smallest open model we know of that does Hindi, English and Hinglish in one file.
At 29.5 MB, Shruti is about 16× smaller than the Hindi-only models we measured it against, and it writes English words in English inside Hindi, the way people type Hinglish. It runs offline on a budget phone: about 1 second per sentence and 80 MB of memory on a Samsung Galaxy F62 with the native engine. In that phone’s browser, the final pass takes about a third as long as the speech. On pure Hindi, the larger Hindi-only models are clearly more accurate.
Of the open models we measured, it makes the fewest errors on Hinglish (MUCS 2021, 3,136 clips), including two Whisper models fine-tuned for Hinglish, and it stays ahead when both sides are written in Roman script.
One small file
Shruti is a single 29.5 MB file: the whole model, 55M parameters, for three ways of speaking. The Hindi-only models in our comparison are 481 and 482 MB, and Whisper-small is 971 MB. A file this size can ship inside an app or download over a phone connection, and it runs with no server at all.
It is built on Cactus Whistle, an open speech model from Cactus Compute. We fine-tuned all 55M of its weights, added 714 Devanagari vocabulary pieces, and trained it on 2,732 hours of openly licensed Indian speech.
How it compares
Word error rate, the share of words a model gets wrong, on the same test sets with the same scorer for every model. Shruti is the smallest file here. It makes the fewest errors on Hinglish (MUCS), it is more accurate than Whisper large-v3 on all three Hindi sets, and it invented no words on 100 silent clips. On pure Hindi, the Hindi-only models Koyal-hi-120M and IndicConformer, about 16× larger, are clearly more accurate. On English, Whisper large-v3 is clearly better, and on Indian English (Svarah) Oriserve’s Whisper-Hindi2Hinglish-Swift is ahead too.
| Model | File size | Monsoon Hindi | FLEURS Hindi | Lahaja (Hindi accents) | MUCS Hinglish | Svarah (Indian English) | LibriSpeech (English) | Words invented on 100 silent clips |
|---|---|---|---|---|---|---|---|---|
| ShrutiUdgam Labs | 29.5 MB (smallest) | 20.46 | 19.92 | 26.28 | 35.21 (best) | 22.47 | 5.95 | 0 (best) |
| Koyal-hi-120MAdalat AI | 481 MB | 11.64 (best) | 9.56 (best) | 18.70 | 57.51 | Hindi only | Hindi only | 0 (best) |
| IndicConformer HindiAI4Bharat | 482 MB | 15.31 | 11.31 | 13.18 (best) | 64.48 | Hindi only | Hindi only | 88 |
| Vosk small (hi + en-in)Alpha Cephei | 139 MB | 35.75 | 18.07 | 26.60 | 61.14 | 38.51 | 30.18 | 1 |
| Whisper-Hindi2Hinglish-SwiftOriserve | 296 MB | 57.16* | n/c | n/c | n/c | 20.16 | 12.60 | 1,676 |
| Whisper-smallOpenAI | 971 MB | 128.25 | 59.42 | 117.5 | – | – | – | – |
| whisper-base-hinglishabhishekgautamm | 290 MB | 65.66* | – | – | 87.05* | 27.11† | 15.53† | 500 |
| whisper-tiny-hinglish-v2abhishekgautamm | 151 MB | 70.60* | – | – | 87.73* | 34.50† | 23.92† | 500 |
| Whisper large-v3OpenAI | 3,087 MB | 33.29 | 28.88 | 31.19‡ | 79.94‡ | 7.39‡ (best) | 2.25‡ (best) | 538 |
Word error rate %, lower is better. Same test sets and the same scorer for every model; Monsoon Hindi with the Hugging Face Hindi leaderboard’s scorer, on the public split. Shruti was run in its on-device engine on a Mac, the other models in the same harness on a GPU or a Mac. Measured 18–19 October 2026.
The best number in each column is in bold. Hindi only: the model does not transcribe English. n/c: not comparable, because the model writes Hindi in Roman letters only. *The model writes Roman script only, so this column mostly measures the script, not the hearing. –: not measured. The silence test uses silent clips only, not noise or music.
†First 1,500 Svarah and first 1,000 LibriSpeech clips; Shruti on the same clips scores 21.35 and 6.36. ‡Whisper large-v3 (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 scores 22.87, 21.35, 5.76 and 45.3. On Monsoon and FLEURS both ran on every clip; in that run Shruti scored 20.14 on FLEURS (19.92 in the table is the earlier engine build).
Scored with both sides converted to Roman script by the same function, MUCS Hinglish is: Shruti 34.32, whisper-base-hinglish 43.45, whisper-tiny-hinglish-v2 44.94. The whisper-*-hinglish model cards report 28.4 and 32.0 on a 250-clip sample with their author’s own converter, which we did not run.
For context, the Hugging Face Hindi leaderboard publishes these Monsoon word error rates: Whisper large-v3 (1.55B parameters) 28.17, Whisper large-v3-turbo 40.79, Voxtral Mini 3B 22.9, Qwen3-ASR-0.6B 18.13 and IndicConformer-600M 8.47. Shruti’s 20.46 is on the public split, and the board’s own number for it may differ by a few points. An official rank needs a submission, which we have not made.
Next to Sarvam’s speech models
Two of Sarvam’s speech-to-text models for Indian languages: Sarvam Edge, which runs on its partners’ devices, and Saaras v4, a paid cloud API. We have not measured either, so this table compares what is published, not accuracy.
| Fact | Shruti | Sarvam Edge speech-to-text | Sarvam Saaras v4 |
|---|---|---|---|
| Availability | Open weights, Apache-2.0 | Closed; commercial partners only | Paid cloud API |
| Size | 29.5 MB file, 55M parameters | About 294 MB (FP16), 74M parameters (Sarvam’s figures) | Server model |
| Internet needed | No: phone, laptop or browser | No, on partner devices | Yes |
| Shown running on | Samsung Galaxy F62 (2021, 6 GB RAM) | Flagship chips (for example Snapdragon 8 Gen 3) in Sarvam’s demos | Servers |
| Languages | Hindi, English, Hinglish | 10 Indian languages | 22 Indian languages and English |
| Published accuracy on shared tests | Yes, on this page | None published | Not measured by us |
Sarvam’s 294 MB is unquantised (16-bit); at 4 bits it would be roughly 37 MB. Saaras v4 is a large cloud model and is likely more accurate than Shruti, and Sarvam Edge is not available to test, so this compares facts, not accuracy.
Sources: Sarvam Edge launch post (14 February 2026) · Sarvam Edge product page · Sarvam model docs (Saaras).
On a budget phone
Samsung Galaxy F62: Exynos 9825, 6 GB of memory, Android 13. A five-year-old phone, launched in 2021. Offline, on 2 threads.
- A sentence of about 7 seconds
- ~1 s
- Including loading the model
- A 60-second recording
- 9.5 s
- 6.3× faster than real time
- Peak memory
- 80 MB
- Over the same runs
42 of 51 transcripts were identical to the Mac’s; the rest differ by a letter or two.
- Browser engine (WebAssembly, one thread, run in Node on a Mac)A 9.6 s clip in 0.45 s, with transcripts identical to the native engine
- Hugging Face Space (a shared 2-core CPU)A 2.8 s clip in 0.35 s
Good fits, and where it falls short
Good fits
- Voice input in apps where the audio should stay on the device.
- Hinglish dictation, notes and search, with English words kept in English.
- Phones with little memory, and places with a weak connection or none.
- Live text from a microphone, with the engine’s streaming calls.
Where it falls short
- Less accurate than larger Hindi-only models on pure Hindi. Koyal-hi-120M and IndicConformer, about 16× larger, make fewer errors on every Hindi test: 9.56% and 11.31% on FLEURS Hindi against Shruti’s 19.92%.
- Weaker on Hindi accents. On Lahaja, IndicConformer makes 13.18% word errors and Shruti 26.28%.
- Hindi comes out in Devanagari only. It does not write Hindi in Roman letters.
- Less accurate than Whisper large-v3 on English: 7.39% against Shruti’s 21.35% on Indian English (Svarah) and 2.25% against 5.76% on LibriSpeech, on the same clips.
- Names it has never heard are spelled by sound. In our own phone test it wrote its name, Shruti, as “सूती” and Udgam Labs as “outgum labs”.
- Far-away or speaker-played audio can throw it. Played from a small computer speaker across a desk, Hinglish was often taken for English; the same clips given as files were transcribed correctly. Speak close to the device.
- Tested on silence, not on noise or music. It invented no words on 100 silent clips.
- Not tested on Windows by us. The engine has a Windows build; we measured on a Mac, Linux, an Android phone and WebAssembly.
- No leaderboard rank. Our Monsoon number is on the public split; we have not made an official submission.
The name
Shruti
श्रुति
Shruti (श्रुति) is Sanskrit for “that which is heard”.
A small model that listens, on your device.
The model
- Parameters
- 55M
- File
- 29.5 MB, one .cact file (W4A8: 4-bit weights, 8-bit activations)
- Built on
- Cactus Whistle by Cactus Compute (Apache-2.0). All 55M weights fine-tuned; 714 Devanagari vocabulary pieces added
- Training data
- 2,732 hours of openly licensed Indian speech (see the credits below)
- Languages
- Hindi (in Devanagari), English, and Hinglish with English words in Latin script
- Input
- 16 kHz mono audio of any length: one pass up to 30 s, longer audio in 30 s windows
- Engine
- Cactus needle (Apache-2.0) on Mac, Linux, Windows, Android, iOS and in the browser (WebAssembly). Measured by us on a Mac (Apple silicon), Linux (x86), an Android phone and WebAssembly; not tested by us on Windows
- Licence
- Apache-2.0 (weights), keeping Cactus’s notice
Use it in your app
Shruti runs in Cactus’s open-source needle engine. Download the engine for your platform and the model file, and pass it a WAV file or 16 kHz audio.
# Command line (Mac, Linux, Windows)
# Engine: huggingface.co/Cactus-Compute/needle3, your platform
# Model: huggingface.co/udgamlabs/Shruti, shruti.cact
./needle --model shruti.cact --audio clip.wav
# for speech that is only English: --audio-language en
# Python
pip install cactus-needle
import needle
shruti = needle.Whistle(weights="shruti.cact")
print(shruti.transcribe("clip.wav")["text"])- Android, iOS and the web: call the engine’s C API.
needle_loadtakes the model file,needle_transcribea clip, andneedle_stream_transcribe_processabout a second of live audio at a time. The demo on this page makes the same calls, from WebAssembly in a Web Worker. - Leave the language unset for Hindi and Hinglish: detection does as well as setting it on Hindi, Hinglish, accented Hindi and US English. Set English when the speech is only English with an Indian accent, which detection often takes for Hindi. On a subset of Svarah: 21.13% word errors with English set, 25.75% detected.
- Keywords, one per line, favour the names, places and product words your users say.
- Cactus’s Python package (cactus-needle) sends anonymous usage counts by default; set
NEEDLE_TELEMETRY=0to turn that off. The command-line engine and the WebAssembly engine on this page make no network requests of their own.
From an AI agent
Shruti also runs as a tool on our MCP server, so Claude Code, Codex, Cursor and other MCP clients can call it. The agent sends a link to the audio; the model’s Hugging Face Space downloads it and returns the transcript. To keep audio on your own device, use the engine above or the demo on this page.
# Claude Code
claude mcp add --transport http udgamlabs https://mcp.udgamlabs.com/mcp
# Codex
codex mcp add udgamlabs --url https://mcp.udgamlabs.com/mcp
# Cursor: .cursor/mcp.json
{ "mcpServers": { "udgamlabs": { "url": "https://mcp.udgamlabs.com/mcp" } } }The tool is shruti_transcribe; the same server has every model that runs as a tool (all tools, and setup for Claude, VS Code and other apps). The model runs on a shared free CPU on Hugging Face, so the first call after a quiet spell waits while it starts.
Credits
Shruti is built on Cactus Whistle and runs in the needle engine, both by Cactus Compute under Apache-2.0. The engine files on this page are Cactus’s WebAssembly build, unmodified, served with its licence and notice. Shruti learned from these openly licensed speech datasets, all of which allow commercial use:
- AI4Bharat IndicVoicesCC-BY-4.0
- ARTPARK-IISc VaaniCC-BY-4.0
- AI4Bharat ShrutilipiCC-BY-4.0
- AI4Bharat NPTELCC-BY-4.0
- AI4Bharat KathbathCC-BY-4.0
- Mozilla Common Voice 17: English, India and South Asia accentsCC0
- LibriSpeechCC-BY-4.0
- MUCS 2021 Hinglish, via dianavdavidson/MUCS-HinglishCC-BY-4.0 on the dataset card; the original is CC-BY-SA-4.0
- AI4Bharat Rural_Women_ASR_v2CC-BY-4.0
- Google FLEURS, Hindi train splitCC-BY-4.0
Weights and links
- Shruti on Hugging Face
The model file (shruti.cact) and model card
- Hugging Face Space
The demo on a server CPU
- Shruti on GitHub
Runnable examples (Python, command line, a web page you can host), checksums and the full comparison
- Cactus needle engines
Prebuilt engines for each platform, and the WebAssembly build
- Cactus Whistle
The speech model Shruti is built on
- needle on GitHub
The engine’s source and documentation
Shruti’s weights are released under Apache-2.0, built on Cactus Whistle (Apache-2.0, Cactus Compute). The model card lists every training dataset and its licence.