A Study of Speech Recognition, Speech Translation, and Speech Summarization of TED English Lectures

Our research focuses on developing an automatic speech recognition system for English lectures, which involves summarizing the content and providing Japanese subtitles. Subtitling the entire audio of an English lecture could hinder comprehension and readability, so a summarization system is desired. By employing the DNN-HMM based speech recognition system, we achieved an 88% word accuracy for recognizing TED lecture speeches. Speech translation results showed a lower BLEU score of approximately 14% compared to text translation. Conversely, speech summarization proved its robustness to speech recognition errors, as the extracted important sentences were almost the same as those in the text summarization process.