Uma arquitetura baseada em aprendizado profundo com mecanismos de atenção para a tradução contínua da língua brasileira de sinais em contextos sem intérpretes

Abstract

In Brazil, the deaf represent about 5% of the population approximately 9.7 million Brazilians. Despite the Brazilian Sign Language (Libras) being recognized as an official language in Brazil, the knowledge and mastery of Libras among the non-deaf is an obstacle, which ends up creating language barriers in accessing basic rights, especially in accessing health services. This has motivated the development of government policies that oblige service providers to provide Libras interpreters to enable access to these services by the deaf community. However, human-based approaches have a high implementation and maintenance cost. From this perspective, it is necessary to develop research and automated methodologies for automatic translation of Libras. Thus in this work, we proposed a methodology for continuous translation of Libras. The proposed methodology does not require any additional hardware, relying entirely on images or image sequences (videos). In addition, a new dataset for Continuous Sign Language Recognition (CSLR) was introduced, containing 10500 videos of 105 distinct sentences in the context of clinical screening. The evaluation experiments achieve a WER of 21.62, while maximum accuracy of about 92.68% for a set of tests with videos and signers never seen by the model during training.