Developing a speech corpus from web news for Myanmar (Burmese) language

Speech corpus is important for statistical model based automatic speech recognition and it reflects the performance of a speech recognizer. Although most of the speech corpora for resource-riched languages such as English are widely available and it can be used easily, there is no Myanmar speech corpus which is freely available for automatic speech recognition (ASR) research since Myanmar is a low resource language. This paper presents the design and development of Myanmar speech corpus for the news domain to be applied to convolutional neural network (CNN)-based Myanmar continuous speech recognition research. The speech corpus consists of 20 hours read speech data collected from online web news and there are 178 speakers (126 females and 52 males). Our speech corpus is evaluated on two test sets: TestSet1 (web data) and TestSet2 (news recording with 10 natives). Using CNN-based model, word error rate (WER) achieves 24.73% on TestSet1 and 22.95% on TestSet2.