开发者在 Hacker News 上发布了名为 RealtimeVAD 的开源工具,旨在解决本地语音助手中的语音结束信号延迟问题 1。该方案使用自适应阈值的能量检测器替代现有的 Silero 模型,通过牺牲一定的误报率来换取更快的响应速度 1。与存在可见语音结束信号滞后的 Silero 模型相比,RealtimeVAD 采用 10 ms 能量检测器配合适应环境噪音的阈值 1。
基准测试显示,该方法在初始阶段比现有方案快约 110 毫秒,修复静音问题后仍保持约 100 毫秒的优势 1。测试基于约 3200 个交互回合进行 1。这一权衡策略是以更多的误报(false positives)换取更低的延迟,并在生成管道中通过快速中止来处理错误 1。
A developer has released RealtimeVAD, an open-source tool on Hacker News designed to reduce the latency of end-of-speech detection in local voice assistants 1. The solution replaces the Silero model with an adaptive energy detector that operates on 10 ms windows and adjusts its threshold based on ambient noise levels 1. By prioritizing speed over accuracy, the system accepts a higher rate of false positives, which are subsequently managed by quick abort mechanisms within the generation pipeline 1.
Benchmark tests conducted across approximately 3,200 interaction rounds indicate that RealtimeVAD delivers a response time advantage of roughly 110 milliseconds compared to existing methods in initial trials 1. After addressing issues related to silence detection, the tool maintained a consistent lead of about 100 milliseconds 1. This trade-off allows for significantly earlier signal transmission regarding speech completion while mitigating the visible lag often associated with the Silero model 1.
评论
还没有评论,欢迎留下第一条。