Korean LVCSR system development for personal assistant service

This paper describes the development of Korean automatic speech recognition (ASR) system for personal assistant service in mobile phone. Methods are presented for Korean ASR word unit selection, building baseline acoustic model (AM), language model (LM) and additional optimization methods for noise robust acoustic modeling, transcription data selection and language modeling using contextual information. We presented effectiveness of proposed method by evaluating it under real use case scenario in mobile environments. The resulting ASR system shows significant performance improvements with these proposed methods and it achieved the performance suitable for consumer product in mobile phone.