Getting two birds with one phone: An acoustic sensor for both speech recognition and medical monitoring

Automatic speech recognition (ASR) in noisy environments requires innovative use of disparate technology to overcome the special demands caused by multiple speakers and minimal signal‐to‐noise ratios (SNRs). Use of nonairborne acoustic sensors for ASR imposes special requirements on speech engines due to the changes in spectral information caused by alternative pickup locations. Specifically, the relative power of voiced and nonvoiced components is often reversed when compared to the relative power of these components collected with conventional microphone technology. Our research entails the evaluation of various ASR sampling configurations in conjunction with different body location points for the physiological sensor. Physiology provides clues to speakers’ stress or cognition. For ASR, our goal is to build a suite of optimal sampling configurations for several strategic body locations (e.g., throat temple, thorax, etc.). The experimental design includes traditional word error rate and subjective task completion components. These experiments were conducted in environments with SNR ranges of 10, 3, 0, and −1 dB. These SNR ranges cover the optimal commercial ASR environment of 10 dB to ranges where commercial ASR systems with conventional microphone technology are completely ineffective. Pilot studies indicate good performance below 0‐dB SNR for a sensor that is throat located.