Reconhecimento de voz utilizando WAV2VEC 2.0 para o português brasileiro

• Main Author: Gris, Lucas Rafael Stefanel
• Publication Date: 2021
• Format: Bachelor thesis
• Language: por
• Source: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Download full: http://repositorio.utfpr.edu.br/jspui/handle/1/29999
• topic: Sistemas de reconhecimento de padrões; Redes neurais (Computação); Reconhecimento automático da voz; Pattern recognition systems; Neural networks (Computer science); Automatic speech recognition; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Deep learning techniques have been shown to be efficient in various tasks, especially in the development of speech recognition systems, that is, systems that aim to transcribe a sentence in audio in a sequence of words. Despite the progress in the area, its development can still be considered a difficult task, especially when there is a lack of data available, as in Brazilian Portuguese. In this sense, this work aims to validate the development of an Automatic Speech Recognition using only open available audio data, from the fine-tuning of the Wav2Vec 2.0 XLSR-53 model pre-trained in many languages, for the Brazilian Portuguese. The final obtained model presents a WER of 11.95%, 13% less than the best open ASR model for Brazilian Portuguese available, which is a promising result in the area. In general, this work validates the use of self-supervising learning techniques, in special, the use of the Wav2vec 2.0 architecture in the development of robust ASRs, even when there is a few available data, and also exposes possible enhancements that can improve even more the obtained result.