Speech recognition is a system that allows for automatic conversion of speech into written text. Such systems typicaly return raw text without any formatting such as capital letters or punctuation symbols. Because such text is unreadable and it also requires a lot of work to edit manually, various methods have been introduced that solve these problems automatically.
Such systems can be based on a variety of methods. However, due to good results they provide, different types of neural networks are mainly used nowdays. As part of the master's thesis, we have implemented a system that uses recurrent neural network to predict punctuation symbols in raw unpunctuated text. We have tried it with different word embeddings such as GloVe, ELMO and BERT. We have also implemented a web service that allows us to easily integrate the system into various other services, such as automatic speech recognition.
|
