An assessment of speech related information contained in GEMS signals

As the essence of communication speech intelligibility, rather than more general speech quality, can be of paramount importance when communications systems operate in high noise environments. This paper considers applications where the acoustic signal is degraded by noise so as to be effectively lost and applications where it is simply not available. With such applications in mind we report experiments to assess the use of non-acoustic general electromagnetic motion sensors (GEMS). Whilst GEMS signals are essentially immune to background noise they are incomprehensible to the human listener. We show that GEMS signals nonetheless contain meaningful speech information within a usable bandwidth in the region of 1 to 2 kHz and report the first comparison of GEMS signals to acoustic signals in the context of automatic speech recognition (ASR). For a small, isolated digit ASR task in a speaker-dependent mode results show word accuracies of 77% are achieved using GEMS signals alone.