Consumer messaging
EVA
Transcribes and summarises voice notes. The first bot of its kind in Latin America.
- Sector
- Consumer messaging
- Stack
- WhatsApp Business APISpeech recognitionNode.jsAudio pipeline
Our own product. Forward a voice note and get back text, or the two sentences that actually mattered.
Voice notes are cheap to send and expensive to receive. Thirty seconds of typing saved becomes four minutes of someone else's time, in a format that cannot be skimmed.
EVA closes that gap. It shipped before the big platforms offered anything comparable in the region, which is why it ended up on television.
Latin American Spanish is not the Spanish in the training data
Accents vary widely, slang moves faster than model releases, and code-switching is normal. Accuracy on real messages was what had to be measured.
Recorded in the worst conditions
Street noise, wind, three people at once, no second take. Whatever arrived is the only sample there will ever be.
Summarising speech is not summarising prose
A voice note circles and puts the one actionable sentence at the end. Methods tuned on written text keep the opening and drop the request.
The unit economics set the constraints
Priced in minutes at a level people in the region will pay. That price dictates the models, the segmentation and the caching.
Tell us where you are starting from. We take it from there.
Start a project