The anatomy of a Turkish speech recognition system

In this paper, we present a Turkish speech recognition system we have developed. We will be giving information about the building of our system and tests we conducted on it. SUVoice voice database, along with METU 1.0, was used in the training of our acoustic models. We performed limited vocabulary and large vocabulary recognition tests. We have shown the performance of modernhidden Markov model based systems for Turkish speech recognition. For simple tasks, we can obtain 1% word error rate, whereas for large vocabulary tests the error rate is higher. This work will constitute a basis for more advanced future works in Turkish speech recognition and contribute to the accumulated knowledge in the field.