Show HN: RealtimeVAD – trading false positives for 110ms earlier end-of-speech (github.com)

A new tool called RealtimeVAD has been introduced to significantly speed up voice activity detection (VAD) in voice assistants, achieving end-of-speech recognition up to 110 ms earlier than the established Silero model. By maintaining the Silero VAD's core capabilities while adding a fast audio level check every 10 ms, RealtimeVAD can detect when a speaker has stopped talking, facilitating quicker responses from voice assistants. With this enhanced accuracy, the tool generates start, pause, and end events, allowing developers to better anticipate user input and reduce delays in processing, ultimately improving the user experience. This innovation is particularly significant for the AI/ML community as it directly addresses a common bottleneck in speech recognition systems—the lag in recognizing when a user has stopped speaking. The tool's design emphasizes both speed and flexibility, performing efficiently on a single CPU thread and requiring minimal dependencies. The ability to process varied audio formats and handle different input types makes it adaptable for diverse applications in real-time speech recognition systems, paving the way for more responsive and intuitive AI-driven interactions in voice technology.