मैं अभी बाहर जा रहा हूं
ನಾನು ಈಗಲೇ ಹೊರಡುತ್ತಿದ್ದೇನೆ.
ASR: 198ms | Translation 7.58s
Zero data leaves your phone
Speak Hear No internet.
HindiTamiltranslated on device
A real-time Indic speech-to-speech pipeline that runs entirely on your Android device. Recognition, translation and speech — all local, all private, all yours.
- ~200 ms
- On-device ASR
- 0
- Network calls
- GPL-3.0
- Open source
Uktam.aiहिन्दीಕನ್ನಡ
ನೀನು ಎಲ್ಲಿ ಹೋಗ್ತಾ ಇದ್ದೀರಾ
तुम कहाँ जा रहे हो?
ASR: 203ms | Translation 7.656s
मैं अभी कॉलेज को जाके वापस शाम पांच बजे को आऊंगा
ನಾನು ಕಾಲೇಜಿಗೆ ಹೋಗಿ ಸಂಜೆ ೫ ಗಂಟೆಗೆ ಹಿಂತಿರುಗುತ್ತೇನೆ.
ASR: 314ms | Translation 11.024s
Four decisions shape the whole app: run locally, use Indic-native models, quantize hard, and never wait on a network.
100% offline & private
No audio and no transcript ever leaves your device. Speech recognition, translation and speech synthesis all run locally on your phone's silicon.
Made for India, by Indian AI
Built strictly on models researched and trained for Indian languages — AI4Bharat and Sarvam AI — so dialects, nuance and grammar land properly.
Custom quantized for mobile
Multi-billion parameter models custom-quantized to GGUF and ONNX for this project, cutting memory and battery cost while holding near-parity accuracy.
No round trip, ever
Speech lands as text in about 200 ms on device. Nothing queues behind an API and nothing degrades when the signal does — dead zones, remote areas and aeroplanes all behave identically.
- 01
Speak
Microphone
live
Talk naturally in Hindi, Kannada, Tamil or Telugu. Audio is captured and streamed straight into the on-device pipeline.
- 02
Recognize
Sherpa-ONNX · IndicConformer
~200 ms
AI4Bharat's IndicConformer transcribes speech in about 200 ms, entirely locally, with no cloud ASR endpoint involved.
- 03
Translate
llama.cpp · Sarvam Translate
7–11 s
Sarvam Translate runs through custom JNI bindings over llama.cpp. A multi-billion parameter model on phone silicon takes a few seconds — and it takes them whether or not you have signal.
- 04
Speak back
Android native TTS
instant
The translated text is spoken aloud through Android's offline text-to-speech engine — a full speech-to-speech loop.
network calls in this pipeline: 0
Four Indic languages today. More on the way.
Speech recognition, translation and text-to-speech in every direction between the supported languages.
हिन्दी
Hindi
hi
ಕನ್ನಡ
Kannada
kn
தமிழ்
Tamil
ta
తెలుగు
Telugu
te
More Indic languages coming soon
0
Indic languages
0
API calls made
0%
On-device compute
Built with modern Android practice and open research from AI4Bharat and Sarvam AI, compiled down to run on ordinary phone hardware.
- Language
- Kotlin
- UI
- Jetpack Compose · Material 3
- Architecture
- MVVM · StateFlow · Coroutines
- ASR
- Sherpa-ONNX · AI4Bharat IndicConformer
- Translation
- llama.cpp (JNI) · Sarvam Translate
- Build
- Gradle KTS · CMake · Android NDK
- Delivery
- Play Asset Delivery
- Licence
- GPL-3.0
Device requirements
- Android
- 14+ (API level 34)
- Memory
- 6 GB RAM recommended
- First run
- ~1–2.5 GB model download
- After setup
- Permanently offline
The three models
One per stage, all of them resident on the device. Open any of them for what it is and why it was chosen.
IndicConformerSpeech recognitionAI4Bharat · ~200 ms
AI4Bharat's Conformer acoustic model, trained on Indian languages rather than adapted to them, which is what keeps recognition honest across accents and code-mixing. It runs through Sherpa-ONNX on the phone's own compute and returns text in roughly 200 milliseconds — fast enough that the transcript appears while you are still talking.
Sarvam TranslateTranslationllama.cpp · 7–11 s
A multi-billion parameter translation model reached through custom JNI bindings over llama.cpp, quantized to GGUF specifically for this project. The app picks the build your hardware can hold: above 6 GB of RAM it loads Q4_K_S for higher accuracy, at or below 6 GB it loads Q2_K for a lighter memory footprint. Either way it takes 7 to 11 seconds, because this is a language model running on a phone rather than in a data centre — and it takes them whether or not you have signal.
Android text-to-speechSpeech synthesisPlatform native · instant
The final step uses the offline voices Android already ships, so there is no fourth model to download and nothing to warm up. It speaks the translation as soon as the text exists, which closes the loop: voice in, voice out, no network touched at any point.
Does Uktam.ai work without an internet connection?
- Yes. After the one-time model download on first launch, Uktam.ai never needs a network again. Speech recognition, translation and speech synthesis all run on the phone, so the app behaves identically in a dead zone, on a flight or in a remote area as it does on wifi.
Which languages does Uktam.ai translate between?
- Hindi, Kannada, Tamil and Telugu, in every direction between them — twelve language pairs. More Indic languages are planned.
Is my voice or transcript ever uploaded?
- No. No audio and no transcript leaves the device. The app makes zero network calls while translating, because every model it uses is running locally on the phone's own silicon.
How much storage does Uktam.ai need?
- Roughly 1 to 2.5 GB, downloaded once on first run depending on which model your hardware gets. After that the app is permanently offline and downloads nothing further.
What Android version and hardware does it need?
- Android 14 or newer (API level 34), with 6 GB of RAM recommended. The app sizes its translation model to the device: over 6 GB it loads a higher-accuracy build, at or below 6 GB it loads a lighter one so it does not run out of memory.
How fast is an offline translation?
- Speech becomes text in about 200 milliseconds. The translation itself takes roughly 7 to 11 seconds, because a multi-billion parameter model is running on phone hardware rather than in a data centre. Speaking the result back is instant.
Is Uktam.ai free, and is the source available?
- Yes to both. Uktam.ai is free on Google Play and the source is published under the GPL-3.0 licence.
Is there an iPhone version?
- Not yet. Uktam.ai is Android-only today; an iOS version is planned but not released.
Translate anywhere.
Even with no signal.
Live on Google Play. Download once, download the models once, then never need a network again.
Get it on Google PlayApp Store — coming soon