A voice message in, a voice answer out

Driving, running, holding a child: nothing to type with and no time to read. A bot that hears and answers aloud covers exactly those, and costs two extra steps.

What is inside

1. Voice from a person — the entry.
2. Transcription turns speech into text.
3. Answer — the model. Its prompt is written for ears: short, no lists, no links, no formatting characters.
4. Speaking the answer turns the text back into sound.
5. Voice answer — the exit, with the file attached.

The prompt matters more than the graph here. An ordinary model answer is full of asterisks, headings and links, and all of it gets read out — the most common reason a voice bot sounds strange while the wiring is perfectly fine.

What to set up for yourself

• The voice and the speech model. Most voice models require a voice, and each has its own list. Leave the field empty and the provider rejects the request.
• The format. mp3 is chosen because every player opens it.
• Answer length. Two or three sentences is deliberate: a minute of synthesised speech takes longer to hear than the same text takes to read.

What to connect

A Telegram bot. Note that a turn here costs more than usual: you pay for transcription, for the model, and for speech. For a conversational bot used several times a day that adds up.