Output details
11 - Computer Science and Informatics
University of East Anglia
Enhancing speech at very low signal-to-noise ratios using non-acoustic reference signals
<22>Speech enhancement is ineffective if the estimates of the interfering noise are poor. Studies into noise estimation have shown conventional methods are only effective at SNRs higher than -5dB. This work is concerned with communication in motor racing where SNRs fall to -15dB. We developed novel methods of noise estimation using telemetric data from the car (engine speed, road speed and throttle) that are substantially better than conventionally derived estimates. When applied to speech enhancement and compared to conventional methods superior speech quality and intelligibility were observed. This work was carried out with a Formula 1 team.