Prosodia

Company

Speech recognition, built in Rosario

Prosodia Sistemas SAS builds speech recognition for rioplatense Spanish that runs on the customer’s own server.

History

4 November 2025Prosodia Sistemas SAS registered in the Registro Provincial de Personas Jurídicas de Santa Fe
May 2026First invoice
Today6 people. 2 paying contact-center sites and 1 court pilot.

Why rioplatense Spanish

The Spanish spoken around the Río de la Plata has features that general models see little of:

  • Voseo: vos tenés, vení, decime, with its own verb endings.
  • “Y” and “ll” pronounced [ʃ] or [ʒ]: yo sounds like “sho”.
  • Aspirated or dropped s: los chicos sounds like “loh chico”.
  • Elision in fast speech: para que becomes pa’ que, and “lo de” becomes one syllable.

We recorded our own corpus in Rosario to cover them, and we measure on audio with these features, not on read speech alone.

Why on-premise

Our customers record hearings and customer calls. Both hold personal data, and Ley 25.326 limits sending it to other countries. Many court networks are closed to the internet altogether.

So the model goes to the audio instead of the audio going to the model. Prosodia Local runs on one GPU server inside the customer’s network, with no outbound connection.

Founders

  • Portrait of Vera Ocampo, CEO of Prosodia.

    Vera Ocampo

    CEO. Licenciada en Lingüística (UNR). 9 years in speech analytics.

  • Portrait of Julián Bersa, CTO of Prosodia.

    Julián Bersa

    CTO. Doctor en Ciencias de la Computación (UNR).

  • Portrait of Nadia Kuriger, Head of Data of Prosodia.

    Nadia Kuriger

    Head of Data. Speech therapist and corpus linguist.

Six people in total. We are hiring for two roles; see careers.

Training compute

Cluster4× NVIDIA H100 80 GB
Use2,100 GPU-hours per month, for training and evaluation

This cluster trains the models. Customer audio never reaches it: production runs on the customer’s hardware.