Venturus, one of Brazil’s leading technology centers, has partnered withthe University of São Paulo (USP) and Swiss startup Data Stand AI to developartificial intelligence models focused on automatic speech recognition (ASR)and text-to-speech (TTS) conversion for spontaneous Brazilian Portuguese.
The project aims to tackle the linguisticdiversity of the Brazilian population—including regional accents, varyinglevels of formality, and prosody—factors that often challenge the accuracy oftraditional ASR and TTS systems trained on formal or read speech data.
“Despite major advances in the field, we stillsee higher error rates in Portuguese compared to languages like English,especially when dealing with spontaneous speech. Many systems also strugglewith socioeconomic bias and have difficulty handling different levels ofeducation, age groups, and regional variation,” explains Frederico Gonçalvesfrom Venturus.
“Our goal with this research is to develop more inclusive models forBrazilian Portuguese, free from these biases, and to make the technology moreeffective in applications like speech transcription, accessibility, and virtualassistants.”
The initiative includes experiencedprofessionals such as Sidney Leal and Frederico Gonçalves from Venturus, aswell as technical support from students and researchers at USP’s Institute ofMathematics and Computer Sciences (ICMC) and the School of Philosophy,Languages and Literature, and Human Sciences (FFLCH). Data Stand AI willcontribute its platform and expertise in data cleaning and annotation.
Over the next two years, the team will developand train robust AI models using real recordings of spontaneous speech. In asecond phase of the project, São Paulo State University (Unesp) is expected tojoin the partnership, expanding the team’s reach and expertise even further.
The project aims to significantly improve theaccuracy of ASR and TTS systems in real-world Brazilian contexts—enabling moreeffective applications in automated services, digital accessibility, education,customer support, and embedded devices. It also represents a key step inadvancing ethically developed AI solutions focused on linguistic inclusion.
