2021/05/13 by Vineet Garg, Garg, Vineet, Wonil Chang +11
Computer Science · #Audio and Speech Processing (eess.AS) #FOS: Computer and information sciences #FOS: Electrical engineering #Human-Computer Interaction (cs.HC) #Machine Learning (cs.LG) #Music and Audio Processing #Sound (cs.SD) #Speech Recognition and Synthesis #Speech and Audio Processing #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2105.06598
openalex publication_date 2021/05/13 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
We present a unified and hardware efficient architecture for two stage voice\ntrigger detection (VTD) and false trigger mitigation (FTM) tasks. Two stage VTD\nsystems of voice assistants can get falsely activated to audio segments\nacoustically similar to the trigger phrase of interest. FTM systems cancel such\nactivations by using post trigger audio context. Traditional FTM systems rely\non automatic speech recognition lattices which are computationally expensive to\nobtain on device. We propose a streaming transformer (TF) encoder architecture,\nwhich progressively processes incoming audio chunks and maintains audio context\nto perform both VTD and FTM tasks using only acoustic features. The proposed\njoint model yields an average 18% relative reduction in false reject rate (FRR)\nfor the VTD task at a given false alarm rate. Moreover, our model suppresses\n95% of the false triggers with an additional one second of post-trigger audio.\nFinally, on-device measurements show 32% reduction in runtime memory and 56%\nreduction in inference time compared to non-streaming version of the model.\n