
RealtimeVAD
Early speech stop signal for voice AI pipelines
Developer ToolsOpen sourceOpen source · verified#voice-assistant#real-time#ai
Listed by the ProductPerch team from Hacker News · not launched here yet
What is RealtimeVAD?
RealtimeVAD is an open source Python library that detects when a speaker stops talking faster than standard voice activity detectors. It wraps around Silero VAD and compares audio levels against room noise every ten milliseconds. When sound drops, it triggers a pause event before the underlying model probability falls. On a close microphone, this pause arrives a median of ninety two milliseconds earlier than Silero, or one hundred thirty milliseconds in fast mode.
The tool supports speculative generation pipelines. Applications can begin processing transcripts, calling language models, or generating voice output at the first pause. If the speaker continues, the system cancels the early work within tens of milliseconds. The library uses a single CPU thread, processes roughly eight milliseconds of audio per second, and requires only numpy and onnxruntime. The Silero model is included, removing the need for PyTorch.
RealtimeVAD provides a drop in replacement for Silero VADIterator and includes a plugin for LiveKit Agents. Benchmarks show consistent gains across quiet speech, steady room noise, and simulated phone lines. The final end signal matches Silero exactly, ensuring no spoken words are cut off.
RealtimeVAD features
- Triggers a pause event before standard VAD models detect silence
- Provides up to 130 milliseconds of lead time over Silero VAD
- Supports speculative generation with fast cancellation logic
- Runs on a single CPU thread using only numpy and onnxruntime
- Includes a drop in replacement for Silero VADIterator
- Ships with a built in LiveKit Agents plugin
Who uses RealtimeVAD and for what
- Building voice assistants that need to respond faster than traditional VAD allows
- Implementing speculative generation pipelines that start LLM calls early
- Reducing latency in live streaming voice chat applications
- Integrating faster silence detection into LiveKit based agent workflows
- Optimizing audio processing for meeting headsets and close microphones
RealtimeVAD FAQ
Is RealtimeVAD free?
Yes, RealtimeVAD is released under the MIT license and is available as open source software on GitHub. The website does not mention any paid tiers or commercial licensing options.
Does RealtimeVAD work on Mac?
The library is written in Python and relies on numpy and onnxruntime, which run on macOS, Windows, and Linux. The website does not specify hardware requirements beyond a standard CPU.
What is RealtimeVAD used for?
It is used to detect when a speaker stops talking faster than standard voice activity detectors. This enables voice assistants and AI pipelines to start processing transcripts or generating responses earlier while retaining the ability to cancel if the user continues speaking.
Is RealtimeVAD open source?
Yes, the project is hosted on GitHub under the KoljaB organization and uses the MIT license. The source code is publicly available for review and modification.
Does RealtimeVAD require PyTorch?
No, the library bundles the necessary Silero model and only requires numpy and onnxruntime as dependencies. The website does not mention support for other machine learning frameworks.
How fast is RealtimeVAD compared to Silero?
In default mode, it triggers a pause signal a median of 92 milliseconds earlier than Silero. Fast mode increases this lead to 130 milliseconds on close microphone input. The website does not provide performance benchmarks for distant microphones or noisy environments beyond the listed tests.
Comments (0)
Comments and upvotes open when this product launches on ProductPerch.
More Developer Tools launches
Monitor Laravel scheduled tasks and queues
Convert Cursor rules to AGENTS.md in the browser. No upload.

Feel iOS haptic patterns on a real iPhone, copy the code
Deploy anywhere from your AI tool.
Python code quality analyzer for developers
eBay listing data for apps and agents