Real time gen ai

Google AI Studio
1

Hi Google team,

I’m using Lyria RealTime and recently noticed a significant change in the audio streaming behavior.

The audio chunks I receive can now be around 2.7 seconds long. Previously, the streaming felt much more continuous.

This results in a noticeable pattern during playback:

audio → ~2–3 seconds gap → audio → ~2–3 seconds gap → audio

I have not changed my client implementation or audio settings.

Could you please confirm:

  1. Is a ~2.7 second audio chunk duration currently expected behavior for Lyria RealTime?
  2. Has anything recently changed on the backend / streaming pipeline that could explain this?
  3. Is there a recommended maximum playback-buffer size for handling this behavior?

I’m specifically wondering whether this is a recent server-side change or regression rather than a client-side issue.

Thanks!

2

I’m also seeing gaps with lyria-realtime-exp. Each decoded PCM chunk contains 2 seconds of 48 kHz PCM16 audio, but chunks arrive every ~2.61 seconds.

My tests show roughly 45–46 seconds of audio per minute.

This reproduced with:

  • google-genai 2.22.0, before and after enabling Tier 1 billing

  • A raw WebSocket using Infinite Crate source-derived settings

  • v1alpha

I also tested Google’s Infinite Crate Mac and got the same gaps/pauses.

Is continuous real-time delivery expected for this model? Are there any project settings or known streaming issues I should check?

3

Hello @Olexandr , @kfiya ,

Thank you for sharing, We are looking into this and will follow up here