LIVE

RealtimeVAD Aims to Reduce Latency in Voice Assistants

🤖 AI Models 1 min read 👁 6

A new tool called RealtimeVAD improves voice assistant responsiveness by detecting when a user stops speaking up to 110ms faster than standard models.

RealtimeVAD Aims to Reduce Latency in Voice Assistants🤖AI Models

Photo: Hacker News Front Page

Key points

  • RealtimeVAD identifies end-of-speech 65-110ms faster than standard Silero VAD.
  • The tool uses a dual-check system to trigger events as soon as audio levels drop.
  • It includes dynamic noise floor tracking to improve performance for quiet speakers.
  • The project is available as a LiveKit plugin for easier integration into voice assistant pipelines.

A new open-source tool, RealtimeVAD, has been introduced to address a persistent bottleneck in voice assistant technology: the delay between a user finishing a sentence and the system recognizing that silence. By augmenting the existing Silero Voice Activity Detector (VAD) with a secondary, high-frequency check, the tool identifies the end of speech significantly faster than traditional methods.

The system operates by performing a check every 10 milliseconds, comparing ambient room noise against the speaker's audio levels. While a standard VAD typically waits for a model to confirm the absence of speech, RealtimeVAD triggers an event the moment audio levels drop. Testing across thousands of voice assistant turns indicates a median latency improvement of approximately 92 to 110 milliseconds compared to using Silero alone.

Designed for developers, the tool is configurable and includes specific features to handle echo cancellation when the assistant is speaking. It also features dynamic noise floor adjustments, allowing it to maintain accuracy even when dealing with quiet speakers or poor-quality audio inputs, such as those found on simulated phone lines.

Source Hacker News Front Page — github.com/KoljaB/RealtimeVAD ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.