Case study: a voice bot that transcribes voice messages

Forward a voice message or a video note and you get text; long text also comes with a short summary. Recognition runs on our server with no third-party services, and recordings are not stored. Russian and English; 20 seconds of speech become text in roughly 7 seconds.

The task

The main pain was video notes. A regular voice message can be turned into text in a third-party service, but video notes usually have no answer. Part of the working questions still arrive exactly that way.

The second requirement was privacy. Voice messages often contain things that should not reach third parties, so sending recordings to an external service was not an option.

What we did

Video notes and voice messages

The bot takes both a regular voice message and a video note, using the audio track of the latter. What comes out is text.

A short summary for long recordings

When the speech is long, the bot gives a short summary alongside the full text. Forwarding a long recording is awkward, and what usually matters is the gist.

Our own server, no third parties

Recognition runs on our server; recordings go to no external API and are not stored. For sensitive conversations that is the whole point.

Speed

20 seconds of speech become text in about 7 seconds — noticeably faster than typing a voice message out by hand.

Two languages

Russian and English. Russian and Georgian is the other case, Kartuli.

Screenshots

A voice message or a video note becomes text; long ones get a short summary too

Voice bot — transcribing voice messages and video notes