Skip to main content
ai-coustics models are available through two LiveKit integrations with different APIs and authentication paths. The official LiveKit plugin is the primary path for speech enhancement. Follow LiveKit’s noise cancellation documentation for installation and configuration. The ai-coustics-maintained plugins provide standalone voice activity detection (VAD) and Tyto audio-quality analysis, billed through ai-coustics. These capabilities are unavailable in the compared official-plugin release.

​
Which plugin for what

This comparison covers official plugin 0.3.2 and ai-coustics-maintained plugin 0.2.0 as of September 18, 2026. The official plugin’s VAD adapter in this release consumes metadata from its accompanying enhancement component. It is not interchangeable with the dedicated VAD(model=...) / new VAD({ model }) API. The ai-coustics-maintained release also includes Processor for existing enhancement integrations; new enhancement projects should use the official plugin.

​
Running both plugins today

The two Python distributions install into the same livekit/plugins/ai_coustics/ directory. They cannot safely coexist in one Python environment. Use separate environments until you have verified that a released package resolves the conflict.
A mixed Python installation can produce an ImportError for Model or Analyzer, or a TypeError for VAD(model=...). Recreate an isolated environment with the single distribution your application needs. Do not use installation order to select an implementation. Node packages use different import names. That distinction does not make their objects interchangeable: each VAD expects metadata from its own audio processing path. The ai-coustics quickstart uses only the ai-coustics-maintained components in its RoomIO chain. Evaluate standalone VAD or Tyto in a separate project while retaining your official enhancement deployment.

​
Authentication and provisioning

Keep all SDK and LiveKit keys on the server. Downloading a model is separate from authorizing a processing session. See authentication.

​
If you use Processor today

Plugin 0.2.0 still exposes Processor. Keep your deployment and lockfile while evaluating official enhancement separately.
  1. Record your baseline. Save the package versions, model ID, enhancement parameters, audio format and authentication configuration.
  2. Configure official enhancement. Follow the official installation and model selection instructions. Its audio_enhancement() / audioEnhancement() factory replaces a manually provisioned enhancement Processor in that environment. Model names do not guarantee identical weights across versions.
  3. Review VAD separately. The dedicated VAD and official adapter have different model and metadata paths. Do not carry the dedicated vad.processor into an official-plugin chain or copy legacy sensitivity values without reviewing their ranges.
  4. Compare and recover. Use the same recordings and downstream providers to compare enhancement, speech boundaries and response latency. Test room disconnect and reconnect. Restore the previous deployment and its lockfile if the new combination fails acceptance.
Do not combine a package change, model change and VAD tuning into one unmeasured rollout. The dedicated VAD migration reference explains the difference between legacy energy thresholds and probability thresholds.

​
Questions

For the ai-coustics-maintained plugin, open an issue in ai-coustics/livekit-plugins. Include package versions, runtime, operating system, model IDs, audio geometry and a redacted error. For the official integration, use LiveKit’s documentation and community.