Gemini Native Audio Model Stops Midway While Speaking

[SaqlainXoas]

Description of the Bug

When using the Gemini native audio model for real-time voice generation, the model occasionally stops outputting audio midway through a sentence.

This behavior is inconsistent — sometimes the audio completes normally, but other times it cuts off abruptly without finishing the response.

---

Model & Configuration

model = "gemini-2.5-flash-preview-native-audio-dialog"

CONFIG = types.LiveConnectConfig(
    response_modalities=["AUDIO"],
    media_resolution=types.MediaResolution.MEDIA_RESOLUTION_LOW,
    # Context window compression (commented out due to causing more frequent stops)
    # context_window_compression=types.ContextWindowCompressionConfig(
    #     trigger_tokens=28000,
    #     sliding_window=types.SlidingWindow(target_tokens=13774),
    # ),
    speech_config=types.SpeechConfig(
        language_code="en-US",
        voice_config=types.VoiceConfig(
            prebuilt_voice_config=types.PrebuiltVoiceConfig(voice_name="puck")
        ),
    ),
    realtime_input_config=types.RealtimeInputConfig(
        automatic_activity_detection=types.AutomaticActivityDetection(disabled=False),
        activity_handling=types.ActivityHandling.NO_INTERRUPTION
    ),
    input_audio_transcription=types.AudioTranscriptionConfig(),
    output_audio_transcription=types.AudioTranscriptionConfig(),
    system_instruction=SYSTEM_INSTRUCTION,
    proactivity=types.ProactivityConfig(proactive_audio=True)
)

Notes:

• Commented out context window compression because it seemed to cause even more frequent stops.
• Output is inconsistent — sometimes fine, sometimes stops after just one sentence in the middle of an interaction.

---

Expected Behavior

The model should complete the entire generated audio without stopping unless explicitly interrupted.

---

Actual Behavior

The audio output randomly stops mid-sentence, requiring a retry or manual continuation.

---

Frequency

• Happens more often after some responses, not usually at the very start.
• Can also occur with short responses.

---

Possible Causes (Not Confirmed)

• API issue (most probably)
• Context window handling
• Streaming bug
• Session resumption issue

---

Request

Could you please confirm if this is a known issue and provide any updates or recommended workarounds?

[yancheesetofer]

This also happens to mine, please prio this stuff

[pariah66]

I am also experiencing this issue.

[itismejy]

Happens to me too

[MuhammadAminGit]

Same thing here and not only on native audio but other models like gemini-2.0-flash-live-001 as well

[itismejy]

I think I kind of solved it, when. you run the code use some headphones. The model is hearing itself talk I think, and interrupting itself

[acetrondi]

I think I kind of solved it, when. you run the code use some headphones. The model is hearing itself talk I think, and interrupting itself

Not true, I'm using headphone also tried it on aistudio but same issue happening getting disconnect after few seconds. Please mark as "Not production ready"!