Reconhecimento de fonemas utilizando redes neurais convolucionais para transcrição fonética automática

• Main Author: Dijkstra, Bauke Alfredo
• Publication Date: 2021
• Format: Master thesis
• Language: por
• Source: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Download full: http://repositorio.utfpr.edu.br/jspui/handle/1/24493
• topic: Sistemas de reconhecimento de padrões; Reconhecimento automático da voz; Fonética acústica; Aprendizado do computador; Redes neurais (Computação); Pattern recognition systems; Automatic speech recognition; Phonetics, Acoustic; Machine learning; Neural networks (Computer science); CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão

Phoneme recognition is the ability to extract features to recognize the sound units of words and transcribe them. Phoneme recognition applications provides assistance towards automatic speech recognition, speaker identification, pronunciation error identification, and emotion recognition. In order to carry out the recognition of the phonemes, a preprocessing stage is initially applied in the audios, called acoustic processing, which allows the characteristics to be extracted, noise and differences between speakers to be reduced. Then, a training and classification stage using machine learning algorithms in order to identify the phonemes. This work aims to develop a technique for automatic recognition of continuous speech phonemes. In the development of this project the tests were performed with phonemes extracted from the audio datasets such as TIMIT Acoustic-Phonetic Continuous Speech Corpus, which is an English-speaking dataset with time-aligned orthographic, phonetic and word transcripts, and Brazilian Portuguese-speaking datsets such as Sid and LaPS Benchmark 16k.The Brazillian Portuguese datasets are only transcribed in orthographic form, so it was required to make changes in these datasets to form the phonemes in regards to the audio recordings.The Praat software was used along with the EasyAlign plugin and a script was developed to format the program’s outputs, aligning the phonemes in time with the frames. In acoustic processing, the Kaldi Speech Recognition Toolkit was applied to extract the MFCC and filter banks. For the training and classification, of the bases cited, a convolutional neural network was implemented in addition to a long short-term memory network using the Pytorch framework. The result obtained on the TIMIT base presented a phoneme error rate in the core test of 18,11% using filter banks and an error rate of 19,04% using MFCC. On the merged Portuguese bases LaPS Benchmark 16k and Sid, an error rate of 24,96% was obtained using filter banks and 25,54% using MFCC.