Personal Voice Activity Detection With Ultra-Short Reference Speech
Longting Xu, Mingjun Zhang, Wenbin Zhang, Tianyi Wang, Jiawei Yin, Yu Miao Gao · 2024
Personal Voice Activity Detection (PVAD) is widely used in applications such as voice assistants. To accurately detect the voice activity of the target speaker, PVAD typically requires pre-registering the target speaker’s speech as a reference. However, the excessively long voice enrollment process tends to reduce user motivation. To address this problem, we explore the possibility that PVAD can maintain good performance even with short reference speech. We propose a PVAD network that supports Ultra-Short reference speech, namely US-PVAD. Unlike traditional methods that rely on pre-trained speaker verification models to extract speaker embeddings, US-PVAD allows the direct input of the original reference speech. Since RNN states can memorize historical information and use it to guide subsequent time steps, we employ a DPRNN-based network and use its RNN states as target speaker embedding. This approach eliminates the need for an external speaker embedding extractor with a large number of parameters. Additionally, the RNN states can be continuously updated during voice activity detection, allowing PVAD to obtain sufficient target speaker feature attributes from ultra-short reference speech. Experimental results show that US-PVAD exhibits better performance when using speech under 2 seconds or even as short as 0.2 seconds as the reference speech.