Skip to main content
The Acoustic Echo Cancellation (AEC) template demonstrates how to use our AEC primitive to filter out speaker echo. Without AEC, if you’re not using headphones, the character’s voice may be fed back into the STT input.
The AEC module works only with the local model and uses CPU processing only.

​
Run the Template

  1. Go to Assets/InworldRuntime/Scenes/Primitives and play the AECTemplate scene. AEC00
  2. When the game starts, play the two example audio clips (Farend and Nearend).
The far-end audio comes from the speaker; the near-end audio is captured by the microphone.
  1. Press Generate to produce the filtered audio.
  2. Then press Play to hear the result.

​
Understanding the Template

​
Structure

  • This demo has only one prefab under InworldController: AEC. It contains InworldAECModule.
  • When InworldController initializes, it calls InitializeAsync() on the AEC module (see Primitives Overview).
  • This creates an AECFactory, which then creates an AECInterface based on the current AECConfig.
AEC01

​
Workflow

Pressing the Generate button invokes AECCanvas.FilterAudio(). It first converts the two audio clips (Farend and Nearend) into AudioChunks, then calls InworldController.AEC.FilterAudio() to generate the filtered audio.
AECCanvas.cs
public void FilterAudio()
{
    AudioChunk farendChunk = WavUtility.GenerateAudioChunk(m_Farend);
    AudioChunk nearendChunk = WavUtility.GenerateAudioChunk(m_Nearend);
    m_FilteredChunk = InworldController.AEC.FilterAudio(nearendChunk, farendChunk);
    if (m_FilteredChunk == null) 
        return;
    if (m_CompleteText)
        m_CompleteText.text = "Audio Generated!";
    if (m_PlayFilteredButton)
        m_PlayFilteredButton.interactable = true;
}

​
Tips for Better AEC Performance

When filtering audio, always use the data taken directly from the speaker output rather than the recorded raw audio. Playback characteristics vary by device, and even small timing or amplitude differences can significantly affect the AEC algorithm’s output. This is especially noticeable on laggy devices: the output audio may be delayed or slightly retimed. That retimed version is exactly what you should use as the far-end reference, not the original raw file.