evidoria

← Back to browse

Good practice Imported

VLibras — Brazil's Neural-AI Libras Avatar Reaching 120,000+ Government Websites

Brazil · Brasília · See the Brazil profile · See the Brasília profile

Evidence: Observational / pre–post Top 25% 73/100 · Ask Evidence Copilot about this practice

Brazil's federal government mandates the neural-AI VLibras avatar—built by UFPB's LAVID lab—on public digital services, translating Portuguese into Brazilian Sign Language in real time. It runs on 120,000+ government sites, with 40 million daily accesses.

120,000+
Government and partner websites carrying the widget (2025)
~40,000,000
Daily accesses (2025)
~3,000,000
Translations per month (2025)
127,000+
Training corpus of Portuguese-Libras sentence pairs (VLibrasBD)
up to 12.73%
Translation quality improvement on BLEU benchmark
12,000+ signs
Sign vocabulary
VLibras — Brazil's Neural-AI Libras Avatar Reaching 120,000+ Government Websites VLibras — Brazil's Neural-AI Libras Avatar Reaching 120,000+ Government Websites

Details

Maturity
Established
Promoter
Ministério da Gestão e da Inovação em Serviços Públicos (MGISP) & LAVID/UFPB
Period
2016–2025
Keywords
digital government, accessibility, sign language, public services

Context

Brazil's federal government requires public-sector digital services to be accessible to Deaf citizens, and since 2015 the Digital Video Applications Lab (LAVID) at the Federal University of Paraíba (UFPB) has developed VLibras, a suite of free, open tools that automatically translates Portuguese text and audio into an animated 3D avatar performing Brazilian Sign Language (Libras).

Objectives

The programme aims to make government websites and documents accessible in real time to Deaf citizens without requiring a human interpreter for every interaction, originally built in partnership with the Ministry of Planning (now the Ministry of Management and Innovation in Public Services) and the Ministry of Human Rights and Citizenship.

Activities

The translation engine evolved from rule-based synthesis to neural machine-translation architectures, including Transformer and ByT5 models, trained on the VLibrasBD corpus of over 127,000 Portuguese-Libras sentence pairs, and the sign vocabulary has grown past 12,000 signs through continuous community-assisted expansion via the crowdsourced WikiLibras platform.

Results

As of 2025 the widget is embedded on more than 120,000 Brazilian government and partner websites, handling roughly 40 million daily accesses and 3 million translations per month, with recent academic evaluation reporting a translation-quality improvement of up to 12.73% on BLEU benchmarks, and in 2024 the project received international recognition from the Vienna-based Zero Project disability-innovation initiative.

Conclusions

Independent academic and user evaluations note that the avatar still struggles to reproduce facial expressions and body language that are grammatically essential in Libras, Brazilian officials themselves position VLibras as a complement to rather than a replacement for human interpreters, and the continuing reliance on crowdsourced correction via WikiLibras reflects a structural scarcity of large annotated Libras datasets shared by most sign-language AI systems.

Implementation
Indicative cost
Medium (€50k–€500k)
Time to results
Long (> 3 years)
Staffing & skills
Digital Video Applications Lab (LAVID) researchers and engineers at the Federal University of Paraíba (UFPB), Ministry of Management and Innovation in Public Services (MGISP), Ministry of Human Rights and Citizenship, WikiLibras crowdsourced community contributors

Conditions for success

  • A federal accessibility mandate requiring public digital services to be usable by Deaf citizens
  • A sustained multi-year partnership between two federal ministries and a university research lab
  • A crowdsourced correction and vocabulary-expansion platform (WikiLibras) to offset the scarcity of annotated Libras training data

Common failure modes

  • The avatar still struggles to reproduce facial expressions and body language that are grammatically essential in Libras
  • Reliance on crowdsourced correction reflects a continuing structural scarcity of large annotated Libras datasets
  • A parallel private Libras avatar used by some municipalities shows the market has not converged on one standard

Commonly funded by

National / regional programmes Own resources / municipal budget

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful