Turkish broadcast news transcription revisited

Abstract

In this study a decade old automatic speech recognition system for Turkish broadcast news transcription is revisited and updated with the latest methods. Recently deep learning using artificial neural networks resulted in significant improvements in speech recognition error rates and became the state-of-the-art. Neural network based acoustic and language models are used as the main components of the speech recognition system built in this paper. For acoustic modeling, deep neural networks are optimized using both cross-entropy and sequence discriminative objective functions. In addition, time-delay neural networks are used for modeling long term dependencies with similar performance to recurrent neural networks. The lowest error rates are obtained using discriminatively trained versions of these models. For the language model a recurrent language model is used. It was observed that the word error rates are approximately halved and fell below 10%.