Energy-efficient speech enhancement processor for user speech recognition in HMD systems

Abstract

Speech recognition has recently come into the spotlight as a User Interface (UI) in the head-mounted display (HMD) system such as a Google glass. However, speech recognition accuracy is significantly degrad-ed in the multi sound source environment, because other unrelated sound sources interfere the user’s voice. In this situation, the speech recognition accuracy can be improved by Speech Extraction (SE) technique, which extracts only user’s voice from the noisy signal. So, SE enables robust speech recognition even in the multi-source environment. Furthermore, Adaptive Noise Cancellation (ANC) reduces the additive channel noise. With the help of this schemes overall speech recognition accuracy can be boosted. For acceleration of the overall speech enhancement algorithm an energy-efficient speech extraction processor is proposed for the robust user speech recognition in the HMD systems. For the low-latency speech extraction, FastSE algorithm is proposed to overcome time-consuming cICA-based user speech selection pro-cess, which results in <2ms SE latency. Moreover, a reinforced-FastSE (RFSE) scheme is proposed to achieve 97.2% accuracy with only 35kB on-chip memory for the low-power HMD applications. Also, reconfigurable matrix operation accelerator (RMAT) is implemented for energy-efficient acceleration of dominant matrix operation in SE. As a result, the proposed SE processor achieves 1.3 times higher speed with 4.24 times small-er memory compared to the state-of-the-art work [4], so that speech recognition in noisy environment be-comes possible for mobile HMD applications.