Vagmi

What Vagmi does

Two modes share the same language pair you configured on first launch:

• Voice translation — tap the record button, speak, and the translation streams onto the screen and plays through the speaker or your headphones.
• Sign reader — point the camera at a sign, name board, menu, or printed text, and get the meaning back in your language.

Everything happens locally on your phone. There is no server round-trip, no recording uploaded, no transcript stored on anyone’s cloud. Vagmi works on a plane, in a metro tunnel, or in a village with no signal.

Voice translation, step by step

1. Tap the big record button on the home screen.
2. Speak naturally — a sentence, a question, whatever.
3. The translation appears on screen as it’s generated and plays in the other language.
4. Tap the swap arrow to flip direction when the other person replies. The languages stay anchored on screen; only the arrow rotates so you always know which way translation is flowing.
5. Tap the record button again to stop.

With headphones connected, Vagmi splits the audio routing intelligently: your voice is captured through the headset mic, the translation you need is delivered privately to your headphones, while translations meant for the person you’re talking to play out through the phone’s loudspeaker. It mimics having an interpreter whispering in your ear.

Sign reader

1. Tap the eye icon in the top-left to switch to sign reader mode.
2. Tap the camera button — the system camera opens.
3. Frame the sign, name board, or printed text and shoot.
4. The result streams into the card below the preview. Tap the speaker icon to hear it spoken.

Choose Transliterate to see the same word in your script (useful for pronouncing place names) or Translate for the meaning. You can flip between the two without recapturing — the last image stays loaded.

Wear headphones if you can

Without headphones, the phone’s loudspeaker plays the translation and the mic re-captures parts of it — you lose fragments of whatever is being said next. Headphones (wired, USB-C, or Bluetooth) keep your mic clear and let you barge in mid-translation to correct or interrupt.

A note on language support

Vagmi works best when the language being translated from has a strong online presence — English, Hindi, Spanish, Mandarin, French, Arabic, Portuguese, Japanese, German and similar. Smaller scripts and low-resource languages can occasionally misfire (wrong script, partial translation, or a literal read-back).

This is a property of the underlying neural models, not a bug we can fix in the app. The language you’re translating into is more forgiving; what trips Vagmi up most often is hearing or reading a low-resource source.

Privacy by design

Vagmi is built so that your conversations never leave your phone. The translation models run on-device. There is no remote ASR server, no cloud translator, no analytics pipeline ingesting transcripts.

The only data that crosses the network is what’s strictly required to run the app: an anonymous installation identifier, anonymous usage counters, and model file downloads from our CDN. Full details in our Privacy Policy.

What’s inside

For the curious, the on-device stack is built from open-source neural models:

• Whisper Medium (OpenAI) for speech recognition — turns your spoken audio into text in the source language.
• NLLB-200 distilled (Meta AI) for text translation — converts source text into the target language. 200 languages covered.
• PaddleOCR PP-OCRv5 (Baidu) for sign reading — extracts text from camera images across Indian and international scripts.
• Kokoro and MMS for text-to-speech — turns the translation back into natural-sounding speech.

All of these run on the phone’s own processor — no accelerator card, no remote inference server. The models are downloaded once during onboarding and reused for every conversation thereafter.