LI CREATIVE TECHNOLOGIES INC — Department of Defense SBIR Phase I: We propose to design and implement a system that combines multiple audio transcription and

Description

We propose to design and implement a system that combines multiple audio transcription and multiple translation tools with natural language processing capabilities, such that information can be automatically extracted from audio files with improved performances. In our approach, speech waveforms of a foreign language are first processed to remove background noise using our noise reduction algorithm developed based on our patented auditory transform theory. The clean speech is then feed forward to multiple speech recognizers to convert to text. A error correction algorithm is then applied to reduce word error rates based on the text and a neural language model. Following that, multiple machine translators are used to translate the text to English. An error correction algorithm with English language model is then applied to make further correction. Finally, a natural language processing unit extracts out the entities, associations, concepts, and themes. Based on recent research results, the proposed approach has the potential to reduce word error rates by 20% and improve the entire system robustness. Our solution will leverage our experience and expertise in noise reduction, speech enhancement, neural network training, robust speech recognition and language model construction.