Recognition of Three Simultaneous Speech Signals Using MFT for a Humanoid

摘要

To achieve a humanoid that can recognize simultaneous speech signals, we integrated a sound source separation and an automatic speech recognition (ASR) system. We designed a spectral feature for the ASR base on missing feature theory (MFT), and presented a method of automatic missing feature mask generation using interchannel leak energy obtained by the sound source separation. We used a humanoid SIG2 with eight microphones, and performed experiments on the recognition of three simultaneous spoken sentences. As a result, three simultaneous speech recognition improved in word corrects.