Measuring the performance of an automatic speech recognition system: The effect of speaker gender and speech register.

Abstract

Speech recognition is an important part of artificial intelligence and has gotten a lot better over the years, but there is still room for improvement. Often the automatic speech recognition systems are not trained equally on male and female voices. Speech recognition systems have many uses, for example it can be used to research language acquisition by evaluating their performance on child directed speech. Child directed speech is a speech register that is used when speaking to children. When the two factors of speaker gender and speech register are combined, what does this mean for the performance of an automatic speech recognition system? In this thesis an answer will be given to the question “What is the effect of speaker gender and speech register on the performance of an automatic speech recognition system?”. The output of an existing automatic speech recognition system was evaluated on accuracy, precision and recall, for a male and female voice using child directed speech and adult directed speech. It was found that speaker gender could positively influence performance if the system was trained on that gender. If that was not the case performance would be more negatively influenced. Furthermore, child directed speech has a positive influence on performance in comparison to adult directed speech.