Sound Source Separation for Plural Passenger Speech Recognition in Smart Mobility System

A novel sound source separation (SSS) method developed for a multi-path automatic speech recognition (ASR) system to support a smart mobility is proposed. This method is able to cope with simultaneous utterances of plural passengers in a car and significantly reduces speech recognition errors, which are caused by interfering speeches of fellow passenger. This method is mainly composed of conventional SSS based on Wiener filter, a novel desired speech detector (DSD) to detect isolated utterances, and a DSD-based post processor to remove the interfering speech. The proposed SSS method makes it possible to recognize each desired speech present in a target direction with high accuracy even though more than one passenger utters simultaneously. The experimental results show that the proposed SSS method reduced residual interfering components after Wiener filter, and significantly improved a speech recognition of ASR with two simultaneous utterances.