Speak in your language.
Send a few dozen bytes.
Be heard on the other side.

The problem

Voice is heavy. The links that survive a disaster are thin.

How it works

Five steps from mouth to ear

1. SpeakHold the talk button. The phone listens for a pause of half a second to know a sentence has ended.
2. RecogniseSraVaani, an open speech model from IISc, turns the sentence into text on the phone, about 20× faster than real time.
3. PackThe text is compressed with a per-language dictionary into a small binary frame, encrypted and signed when phones are paired.
4. SendThe frame goes over Wi-Fi, then Wi-Fi Direct, then Bluetooth LE, whichever works. A LoRa bridge extends it to kilometres.
5. Speak againThe other phone voices the text with an Indian TTS model. Alerts jump the queue and play at full volume.

System diagrams

System block diagram: one phone, both directions Full deployment: phones, LoRa relays, control room Sender flow Receiver flow Link fallback: Wi-Fi → Wi-Fi Direct → BLE → LoRa Latency sequence: said → heard

What the prototype does today

Features, with honest status

Each item was checked against the source code on 29 Sep 2026. Built works in the app now. Partial exists with gaps. Planned is not started.

Real screenshots

The app on a vivo I2202 (Snapdragon 870)

Captured from the running prototype. Tap any screen to see it full size.

Talk screen with a flood alert SOS: location and one-tap templates First run: pick your language History in Tamil and Hindi Metrics: latency, memory, thermal Whole interface in Hindi Built on Indian open AI QR pairing for encrypted links Settings: text size, theme, speech rate Light theme

Control-room gateway dashboard, running against the phone: live feed, latency, per-language counts, broadcast and CAP alert ingestion

See it working

Demo video and phone recordings

Measured on real phones

Prototype numbers

From the app's own logging on a Snapdragon 870 (vivo I2202) and a Snapdragon 8 Gen 3 (OnePlus 12), unless the note says otherwise.

A hot phone can run 2 to 4 times slower. The Metrics screen shows thermal state beside every number, so results can be compared fairly.

Open source, end to end

Models and technology

For reviewers

Install and test the prototype

About 15 minutes the first time. You need an Android phone (8.0 or newer, 64-bit, 6 GB RAM or more recommended, about 2 GB free) and, for the speech models, a computer with a USB cable.

1. Install the app

   On the phone, open this page and tap Download APK. Open the file and allow Install unknown apps for your browser when Android asks. This is a debug build signed with a development key, so Play Protect may warn you; choose Install anyway.

   Open iTantra once and grant the microphone and notification permissions.

2. Load the speech models (1.2 GB, once)

   The models are too large to put inside the APK, so they are copied to the phone over USB. Turn on Developer options → USB debugging, connect the phone, and accept the prompt. You also need Android platform-tools (adb).

   macOS / Linux Windows Manual

   curl -fsSL https://itantra-106.pages.dev/install-models.sh -o install-models.sh
   bash install-models.sh

   Copy

   The scripts check every file's checksum and skip files already downloaded, so you can safely run them again. Read them first: install-models.sh · install-models.ps1.

3. Check the models

   Open iTantra → Settings → Models. Every row should show as present. Without the models the app still opens and you can explore every screen, but it cannot recognise or speak.

4. Talk between two phones

   Install on two phones and put both on the same Wi-Fi (a phone hotspot works). They find each other automatically within a few seconds; the link type shows at the top. With no Wi-Fi they try Wi-Fi Direct, then Bluetooth.

   Only one phone? Run the control-room gateway from the source code (tools/gateway) on a laptop on the same Wi-Fi and send messages from its dashboard.

5. Things to try

   • Pick your language, hold the orange button, say a sentence, release. Watch the text appear and the other phone speak it.
   • Turn on Alert and send "flood water is rising". The other phone plays it on the alarm channel at full volume and vibrates.
   • Tap SOS and choose a template. The location is spoken on the other phone.
   • Switch to Call for hands-free conversation.
   • Open Metrics to see latency, real-time factor, memory and thermal state.

What comes next

Roadmap

1. Working prototype10-language speech in and out, alerts, SOS, three links, binary frames, encryption, metrics, gateway, demo video.
2. Idea round30-clip accuracy per language, speech intelligibility test, end-to-end latency p50/p95, a 4 GB phone test.
3. Before the finaleSmall fast voices for hi/gu/or/en, smaller recogniser for low-end phones, LoRa bridge on real hardware, two-phone field test.
4. FinaleLive multi-hop LoRa demo in two languages, phone relaying, place names for SOS, listener study.
5. AfterOn-device translation, the sender's own voice, live NDMA SACHET alerts, pilot with a district disaster team.

Credits and references

Built on Indian open AI research