Anti-High Decibel Keyword Spotting System Based on Flexible Sensor

Since speech is considered as one of the most important biological signals, keyword spotting (KWS) system enjoys a high level of acknowledgement when it comes to realizing human-human and human-computer interaction. Acoustic microphones are mostly utilized to collect airborne sound waves, but in factories and other noisy venues, the quality of speech collected by conventional method is far less satisfactory, causing a high false alarm rate. To deal with this problem, a novel embedded KWS system is proposed, where a flexible sensor to capture real-time vibration speech signals from the jaw is exploited. And a KWS algorithm based on non-acoustic signals is designed and transplanted into an embedded system. Compared with an acoustic microphone under 90dB conditions, the system has achieved an excellent anti-noise effect and a desired audio quality, which can be demonstrated by the experimental results, among which the average perceptual evaluation of speech quality(PESQ) and average short-time objective intelligibility(STOI) are 3.31 and 0.76, respectively. On top of that, the average accuracy of KWS is 83.53% under 90dB conditions. Its stable performance and resistance to noise make the system a promising tool for next-generation KWS and interaction systems.