mmMIC: Multi-modal Speech Recognition based on mmWave Radar
Long Fan, Lei Xie, Xinran Lu, Yi Li, Chuyu Wang, Sanglu Lu
Abstract
With the proliferation of voice assistants, microphone-based speech recognition technology usually cannot achieve good performance in the situation of multiple sound sources and ambient noises. In this paper, we propose a novel mmWave-based solution to perform speech recognition to tackle the issues of multiple sound sources and ambient noises, by precisely extracting the multi-modal features from lip motion and vocal-cords vibration from the single channel of mmWave. We propose a difference-based method for feature extraction of lip motion to suppress the dynamic interference from body motion and head motion. We propose a speech detection method based on cross-validation of lip motion and vocal-cords vibration so as to avoid wasting computing resources on nonspeaking activities. We propose a multi-modal fusion framework for speech recognition by fusing the signal features from lip motion and vocal-cords vibration with the attention mechanism. We implemented a prototype system and evaluated the performance in real test-beds. Experiment results show that the average speech recognition accuracy is 92.8% in realistic environments.