Deep neural network based phonological feature extraction for Bengali continuous speech

Automatic Speech Attribute Transcription framework is a recently proposed paradigm for detection based bottom-up speech recognition. Speech signal contains a large set of related information, known as speech attributes that include a set of fundamental speech sound with their linguistic identification that is phonological features. In this study, a bank of Deep Neural Network based attribute detectors has been applied to Bengali continuous speech corpus to detect phonological features of Bengali language. 89.17% of average attribute detection accuracy was achieved for Bengali continuous speech. This experiment was repeated using TIMIT speech corpus to ensure the system robustness, and the average detection accuracy was 89.40%.