Transcribing every word whoever speaks

Automatic speech recognition is considered to be a mature technology that can accurately recognize sentences that are correctly and formally uttered. However, the recognition of casual speech is still a challenging task since speakers and speaking styles inevitably vary greatly in such situations. To tackle this, we developed a unified statistical model for both acoustic and language representations, and enhanced it by using a multi-layered algorithm that extracts recognizer-friendly representations from the observed speech by combining simple processing layers. Using these technologies, we are investigating a robust speech recognizer that can transcribe every word that anyone speaks. The recognizer will help to promote a natural interaction between humans and machines.