Skip to content
AlphaMergeProduct brief 05 — NativeChatPhase 1 live · Android in testing
Product brief 05·Phase 1 — live on Telegram

NativeChat — a voice note in any language, back in yours.

Send a voice message in any language. Get back what it says in yours, as text and as a voice note. No typing, no choosing a language per message, and no need to know what language the recording was in before you play it.

Live

Running as @NativeChatterBot on Telegram. That is phase 1 and it is deliberately the smallest surface that proves the idea. Phase 2 is an Android app in private testing, and it took a different route: the same three steps, but all of them on the phone.

2

What it removes

Translation tools mostly assume you already know what you are dealing with: paste the text, pick the source language, read the result. A voice note from a supplier, a relative or a driver satisfies none of those assumptions — it is audio, in a language you may not recognise, arriving in a thread you are already in. The product is the removal of those steps, not the translation itself.

01 You never pick a source language. The recording is identified, not declared.
02 You never type. Speech goes in, speech and text come back.
03 The sender does nothing differently. They send a voice note as they always would.
04 It answers where the conversation already is, rather than asking anyone to move.
3

The pipeline

Three hops

Speech to text, translation, then text to speech — in the bot each a managed service, on the phone all three local, and none of them the interesting part. What makes it a product rather than three API calls is that the language is decided once and never asked about again.

01 · Speech to text

The recording is transcribed in whatever language it happens to be in.

02 · Translation

The transcript is translated into the language you set once, not per message.

03 · Text to speech

The result comes back as a voice note as well as text, so it can be listened to.

Every provider can be faked locally. The whole Telegram flow runs end to end with nothing billed and no credentials present, which is what keeps the loop fast and stops a cloud account being a prerequisite for working on it.

4

Status

Phase 1 is a dockerized bot, live and usable now. Phase 2 is an Android app in private testing: speech recognition, translation and noise removal all run on the handset, so a recording never leaves it. It listens for its owner's voice, floats over whatever app you are in, and picks up incoming WhatsApp voice notes on its own. iOS has not been started, store billing is not wired up, and the compliance work that comes with handling voice is still ahead of us.

Phase 1LiveTelegram bot, dockerized
Phase 2In testingAndroid app, fully on-device
Planned—iOS · store billing · compliance

You can try it before reading another word.

Send it a voice note in any language. That is the whole demonstration.