Gemini Live Audio Regression (Post-March 9, 2026 Update)

Gemini API
1

Severe audio degradation observed on both Lyra and Pegasus voices, likely due to overly aggressive bitrate reduction/compression.
Symptoms:

  1. Harsh, “glass-scratching” sibilance distortion on fricatives (“S” and “Sh” sounds).
  2. Audio tail clipping (word endings are dropping out or muffled).
  3. Severe loss of prosody/dynamic range—Lyra’s voice has lost its natural cheerfulness and sounds artificially flattened.

Request: Please review the codec or acoustic model changes made on/around March 9th and restore the audio fidelity and vocal warmth present in the February 2026 build.

Devices: Android 16, 15 and 14, on Galaxy A54, A56, Google Pixels 9 and 10a.
Environment: highspeed network, over 150 MBps (no a device, nor the network connectivity is a culprit)

P.S. I posted it on Gemini App Feedback. Since the problem is pretty bad, I’m posting it here.

#GeminiLive
#Regression

2

Update: A/B Testing confirms it is an App/Codec issue, not the core model.
​To isolate the problem, I just tested the audio output using the streaming mode in Google AI Studio. The TTS audio there remains high-fidelity, warm, and completely lacks the harsh sibilance and clipping I’m experiencing in the mobile app.
​This confirms that the underlying acoustic model for Lyra/Pegasus (those voices are the same if the Studio settings left untouched) is still intact. The severe degradation seems strictly localized to the Gemini Live mobile app’s specific streaming codec or an aggressive bitrate compression introduced during the recent rollout.
​Hopefully, this helps narrow down the routing issue for the mobile app!