Traverba: Live Translator

My mother speaks Cantonese. Our domestic helper speaks Tagalog. Every day — meals, schedules, emergencies — they need to communicate, and neither speaks the other's language.

They tried Google Translate. Type, wait for the server, squint at the screen. For two people who just want to talk naturally, it's painful. And when the WiFi drops, it stops working entirely.

I'm a solo developer from Hong Kong. I watched this for months and thought: what if two phones could just translate a conversation in real-time, with no internet at all?

So I built Traverba.

## What it does

Everything runs on-device. No cloud. No data leaves your phone.

- Voice — speak naturally, see the translation instantly

- Camera — point at a menu, sign, or medicine label, get an overlay translation

- Screen — translate any text on your screen without switching apps

- Text — type or paste, 140+ languages

- Bluetooth conversation — multiple phones pair over Bluetooth, each runs its own speech recognition and translation locally, only the translated text crosses the link. No WiFi, no hotspot. Works with groups too.

Cantonese is a first-class language — not buried under "Chinese (Traditional)" like every other translator.

## The mum test

The first time my mum and our helper used the Bluetooth mode, my mum laughed and said "finally I can tell her exactly what I mean."

That was the moment I knew it was worth shipping.

## Who it's for

I built it for my family, but the problem is bigger than I expected:

- Hong Kong has 400,000+ domestic helpers, mostly from the Philippines and Indonesia

- Millions of families across Asia have the same language gap at home

- Travelers hit the same wall daily — the places you need a translator most are the places with the worst connectivity

- 85 million Cantonese speakers worldwide, ignored by every major translator

## How I built it

2 months. Solo. Claude Code as my coding agent.

Flutter for the cross-platform app. Native Kotlin (Android) and Swift (iOS) for the ML-heavy parts. Three speech recognition models — Parakeet (bundled), Whisper.cpp (Metal/Vulkan), Qwen3-ASR (CoreML/ONNX) — because no single model covers all languages well. PaddleOCR v5 for camera and screen text recognition.

The hardest part was fitting ~3.5GB of ML models into memory on a mid-range phone without crashing. Bluetooth Low Energy has tiny bandwidth, so each phone does all the processing locally and only sends text across the link.

## Business model

- Free: voice, camera, screen, text translation — all offline, no limits

- Subscription: $0.49–$1.99/mo (regional pricing) for group Bluetooth conversations and premium models

- Ad-supported coins: watch an ad, earn coins for premium features. You can use everything free forever if you're willing to watch a few ads.

The on-device architecture means my marginal cost per user is basically zero. Even users on $0.49/mo in Southeast Asia are profitable.

## Numbers so far

Just launched. Early days. Working on organic growth with a tiny HK$1,500 (~$190 USD) marketing budget. Most of my strategy is ASO, content, and community — not paid ads.

## What I'd love feedback on

- Does the Bluetooth conversation angle resonate, or does it sound too niche?

- Any suggestions for reaching the domestic helper market in Hong Kong/Asia?

- Fellow solo devs — how do you handle on-device ML memory management?

Try it: https://www.traverba.com

Happy to answer any questions about the tech, the market, or the journey.