All articles
Hackathon1st prize

A WhatsApp bot you can talk to

Second edition, second win: in 48 hours, team Codeur Pro built a voice bot that listens, understands and answers out loud.

by Vidhal Elame
EventHackathon Bafoka 2025 · 2nd edition
DateNov 2025
TeamCodeur Pro
Result1st prize · $1,000

The challenge

Build a voice WhatsApp bot able to interpret, understand and speak back: the theme set by the organisers, ActivSpaces, with the support of GFA Consulting Group and GIZ.

Three months earlier, at the first edition, we had won with BAFOKA-Market, a marketplace inside WhatsApp. This time the subject went one step further: no text at all.

Why voice

A bot that writes assumes the person on the other side reads and types without effort. That is not always true. Many people are more at ease speaking than writing, and on WhatsApp the voice note is already a reflex: press, speak, release.

A voice bot meets those people in their own way of doing things. It does not ask them to learn a menu or type a keyword. They ask their question the way they would ask a person.

What we built

A bot that works in French and English, around three steps:

  • fluid voice interpretation;
  • contextual understanding;
  • natural speech in the answer.

The goal: voice AI and digital accessibility for people who would rather speak than type.

The journey of a voice note

Here is, in broad strokes, what happens between the moment someone speaks and the moment the bot answers.

  1. WhatsApp receives the voice note and hands it to the bot.
  2. A Python service, with Flask, fetches the audio file.
  3. Whisper transcribes it into text. It recognises the words, and the language too: the same person can speak French one day and English the next, with nothing to set.
  4. Botpress takes the text and holds the conversation. It knows what was said just before, which lets it answer "and the price?" without asking what we are talking about.
  5. The answer is turned into speech, in the language of the question.
  6. The bot sends it back on WhatsApp, as a voice note.

None of these steps is magic on its own. The work is chaining them so the person never feels they are waiting for a machine.

What makes voice hard

Delay. Every step adds time. In writing, three seconds go unnoticed. In speech, three seconds of silence feel like nothing works. So answers must be short, and receipt acknowledged at once.

Noise. A voice note is recorded in the street, at the market, in a taxi. Transcription is never perfect. The bot must be able to say "I did not catch that, could you repeat?" rather than answer the wrong question with confidence.

Two languages. Recognising French or English is not enough: the answer must come back in the same language, in a voice that sounds right in that language.

Memory. A spoken conversation is full of things left unsaid. Without the context of the previous messages, the bot asks for everything again each time, and the person gives up.

The podium

1st prize: Codeur Pro. 2nd: Team-Z. 3rd: Gravity Falls.

Thanks

To the jury for their structured feedback, which already pushes us further.

Two editions, two first places, and one common thread: start from the tool people already have in their pocket. Ten days later we were at the Genuka Hackathon with GestAvi.