Due to the advantages of device ubiquity, natural interaction and privacy preservation, acoustic-based gesture input has received widespread attention. Researchers have proposed various techniques for different applications. However, the existing work has shortcomings of heavy data-collection overhead, non-continuous input, and performance degradation in crossuser scenarios. To overcome these shortcomings, we propose UltraWrite, an acoustic-based gesture input system that only needs extremely low data-collection overhead, supports continuous input, and achieves high cross-user recognition accuracy. The key idea of our solution is to synthesize training data of continuous gestures from isolated ones, build a lightweight continuous gesture recognition model based on connectionist temporal classification (CTC) mechanism, and design a novel decoupled model training strategy to improve its cross-user recognition capability. We have implemented prototype systems on commercial devices and conducted comprehensive experiments to evaluate their performance. The results show that UltraWrite achieves an average top-1 word accuracy of 99.3% and top-1 word error rate of 0.34%. In addition, we have also evaluated UltraWrite's robustness to the sensing distance, angle, background noise, and device. The results reveal that UltraWrite possesses strong robustness to these factors.

Due to the advantages of acoustic sensing such as device ubiquity, hands-free interaction and privacy security, acoustic-based gesture input techniques have gained extensive at-tention and many excellent works have been proposed. However, these works have the following shortcomings: high cost of system construction, non-continuous input, and degraded performance in cross-user scenarios. To overcome the above shortcomings, we propose UltraWrite, a continuous gesture input system that needs rather low system construction cost, supports continuous input, and achieves high cross-user recognition performance. The key idea of our solution is to synthesize the data of continuous gestures from isolated ones, and build an end-to-end continuous gesture recognition model by introducing the connectionist temporal classification (CTC) mechanism widely used in natural language processing. We have implemented the system on a commercial tablet and conducted comprehensive experiments to evaluate its performance. The results demonstrate that UltraWrite achieves an average word accuracy of 99.3% and word error rate of 0.34% when considering only the first output candidate word. In addition, we have also evaluated the system's robustness to background noise, sensing distance and angle. The results reveal that UltraWrite displays strong robustness to these factors.