Publicação
Automatic speech recognition for European Portuguese
| Resumo: | The process of Automatic Speech Recognition (ASR) opens doors to a vast amount of possible improvements in customer experience. The use of this type of technology has increased significantly in recent years, this change being the result of the recent evolution in ASR systems. The opportunities to use ASR are vast, covering several areas, such as medical, industrial, business, among others. We must emphasize the use of these voice recognition systems in telecommunications companies, namely, in the automation of consumer assistance operators, allowing the service to be routed to specialized operators automatically through the detection of matters to be dealt with through recognition of the spoken utterances. In recent years, we have seen big technological breakthrough in ASR, achieving unprecedented accuracy results that are comparable to humans. We are also seeing a move from what is known as the Traditional approach of ASR systems, based on Hidden Markov Models (HMM), to the newer End-to-End ASR systems that obtain benefits from the use of deep neural networks (DNNs), large amounts of data and process parallelization. The literature review showed us that the focus of this previous work was almost exclusively for the English and Chinese languages, with little effort being made in the development of other languages, as it is the case with Portuguese. In the research carried out, we did not find a model for the European Portuguese (EP) dialect that is freely available for general use. Focused on this problem, this work describes the development of a End-to-End ASR system for EP. To achieve this goal, a set of procedures was followed that allowed us to present the concepts, characteristics and all the steps inherent to the construction of these types of systems. Furthermore, since the transcribed speech needed to accomplish our goal is very limited for EP, we also describe the process of collecting and formatting data from a variety of different sources, most of them freely available to the public. To further try and improve our results, a variety of different data augmentation techniques were implemented and tested. The obtained models are based on a PyTorch implementation of the Deep Speech 2 model. Our best model achieved an Word Error Rate (WER) of 40.5%, in our main test corpus, achieving slightly better results to those obtained by commercial systems on the same data. Around 150 hours of transcribed EP was collected, so that it can be used to train other ASR systems or models in different areas of investigation. We gathered a series of interesting results on the use of different batch size values as well as the improvements provided by the use of a large variety of data augmentation techniques. Nevertheless, the ASR theme is vast and there is still a variety of different methods and interesting concepts that we could research in order to seek an improvement of the achieved results. |
|---|---|
| Autores principais: | Campinho, Adriano Vaz de Carvalho |
| Assunto: | Automatic speech recognition European Portuguese End-to-end learning Data collection Reconhecimento automático de fala Português Europeu Aprendizagem ponta a ponta Recolha de dados |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | The process of Automatic Speech Recognition (ASR) opens doors to a vast amount of possible improvements in customer experience. The use of this type of technology has increased significantly in recent years, this change being the result of the recent evolution in ASR systems. The opportunities to use ASR are vast, covering several areas, such as medical, industrial, business, among others. We must emphasize the use of these voice recognition systems in telecommunications companies, namely, in the automation of consumer assistance operators, allowing the service to be routed to specialized operators automatically through the detection of matters to be dealt with through recognition of the spoken utterances. In recent years, we have seen big technological breakthrough in ASR, achieving unprecedented accuracy results that are comparable to humans. We are also seeing a move from what is known as the Traditional approach of ASR systems, based on Hidden Markov Models (HMM), to the newer End-to-End ASR systems that obtain benefits from the use of deep neural networks (DNNs), large amounts of data and process parallelization. The literature review showed us that the focus of this previous work was almost exclusively for the English and Chinese languages, with little effort being made in the development of other languages, as it is the case with Portuguese. In the research carried out, we did not find a model for the European Portuguese (EP) dialect that is freely available for general use. Focused on this problem, this work describes the development of a End-to-End ASR system for EP. To achieve this goal, a set of procedures was followed that allowed us to present the concepts, characteristics and all the steps inherent to the construction of these types of systems. Furthermore, since the transcribed speech needed to accomplish our goal is very limited for EP, we also describe the process of collecting and formatting data from a variety of different sources, most of them freely available to the public. To further try and improve our results, a variety of different data augmentation techniques were implemented and tested. The obtained models are based on a PyTorch implementation of the Deep Speech 2 model. Our best model achieved an Word Error Rate (WER) of 40.5%, in our main test corpus, achieving slightly better results to those obtained by commercial systems on the same data. Around 150 hours of transcribed EP was collected, so that it can be used to train other ASR systems or models in different areas of investigation. We gathered a series of interesting results on the use of different batch size values as well as the improvements provided by the use of a large variety of data augmentation techniques. Nevertheless, the ASR theme is vast and there is still a variety of different methods and interesting concepts that we could research in order to seek an improvement of the achieved results. |
|---|