I want to use the new

gemini-2.5-flash-native-audio-preview-12-2025

live model which livekit supports. The normal version works fine but when I try to use it with a 3rd party tts provider e.g elevenlabs, I get the following error:

ConnectionClosedError: received 1007 
(invalid frame payload data) Cannot 
extract voices from a non-audio request.;
then sent 1007 (invalid frame payload 
data) Cannot extract voices from a 
non-audio request.

I use the model as follows:

return google.realtime.RealtimeModel(
        model="gemini-2.5-flash-native-audio-preview-12-2025",
        modalities=[types.Modality.TEXT],
        # voice="Sulafat",
    )

            session = agents.AgentSession(
                llm=llm,
                stt=stt,
                tts=tts,
                vad=vad,
            )

Thanks for reaching out, Muhammad. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

For both STT/VAD, I have this:

return NotGiven()

Here are all the logs I could find:

ConnectionClosedError: received 1007 
(invalid frame payload data) Cannot 
extract voices from a non-audio request.;
then sent 1007 (invalid frame payload 
data) Cannot extract voices from a 
non-audio request.

APIConnectionError: Failed to connect to 
Gemini Live (body=None, retryable=True)
78755 12/19/2025, 7:17:53 PM    ERROR [livekit.agents]   Error in _realtime_reply_task

RealtimeError: generate_reply timed out 
waiting for generation_created event.

Thanks — I see you return

NotGiven()

for STT and VAD and you shared the logs showing 1007 + APIConnectionError + a generation timeout. I’m responding to those errors and your setup. Next steps to isolate and fix this: 1. Confirm the model is explicitly set to text-only • Show me the exact import line you use for

Modality

. It must be the same enum the plugin expects (for example

from google.generativeai import types as genai_types

then

modalities=[genai_types.Modality.TEXT]

). A mismatched enum can let the model still expect audio. 1. Ensure no audio/voice flags are sent anywhere • Search your codebase for any

voice=

,

audio=

, or

modalities

uses elsewhere. Even a stray

voice

or an audio flag will cause the model or LiveKit to expect audio frames. 1. Verify you are not sending any binary/audio frames • Even if STT/VAD return

NotGiven()

, confirm your application is not sending microphone binary frames or previously-buffered audio to the Gemini realtime connection. Log outgoing frames right before sending. 1. Check connectivity / retry behavior • The APIConnectionError indicates intermittent failure to reach Gemini Live; ensure network connectivity and that calls are retried. That connectivity issue can cause timeouts and generation failures. 1. Enable debug logging and capture payloads • Turn on debug logging (

logging.basicConfig(level=logging.DEBUG)

) and capture the exact request payloads and frame types sent to Gemini. If you can paste a single sample request payload (redact secrets), I’ll inspect it. Quick checks you can run now • Confirm

modalities

import line (paste it here). • Grep for

voice=

across your repo and paste any hits. • Share one debug log snippet showing the outgoing request/frame before the 1007 error. If you send the three small items above, I’ll point to the exact field causing Gemini to treat the request as non-audio and show the minimal code change to fix it.

_from_ google.genai _import_ types

There's nothing with

voice=

Here's all the logs I can find:

78755 12/19/2025, 7:17:47 PM    DEBUG [livekit.plugins.google] connecting to Gemini Realtime API...
78755 12/19/2025, 7:17:48 PM    ERROR [livekit.plugins.google] Gemini Realtime API error: received 1007 (invalid frame payload data) Cannot extract voices from a non-audio request.; then sent 1007 (invalid frame payload data) Cannot extract voices from a non-audio request

Thanks — I’m responding to the logs you pasted and your import line (

from google.genai import types

). What I see • Your import is from

google.genai

not

google.generativeai

(the expected enum provider). The Gemini realtime model still appears to be treating the connection as audio-capable and rejecting a non-audio request, which produces the 1007 error. Next steps (try these in order) 1. Use the correct Modality enum • Replace

from google.genai import types

with the plugin/SDK enum used by the RealtimeModel, for example:

from google.generativeai import types as genai_types
  llm = google.realtime.RealtimeModel(
      model="gemini-2.5-flash-native-audio-preview-12-2025",
      modalities=[genai_types.Modality.TEXT],
  )

• Confirm

modalities

is exactly

Modality.TEXT

from that import. 1. Verify nothing else signals audio • Grep your project for

voice=

,

audio=

,

modalities=

or other Gemini-related flags and ensure no other module or default sets an audio modality. • Ensure your AgentSession

tts

is the external ElevenLabs instance (the model should only return text to that session). 1. Test a minimal repro • Run a tiny script that only creates the

RealtimeModel

with

modalities=[Modality.TEXT]

and opens a session (no STT/VAD, no mic frames). If it connects without 1007, the issue is elsewhere in your stack. 1. Capture debug frames if still failing • Enable debug logging and paste the debug lines around the connection (you already have a useful line). If failure persists, share the next block of debug log that shows the initial request/handshake frames (redact keys). One quick question to confirm: after you change the import to

google.generativeai.types

(or equivalent the plugin expects), can you re-run the minimal repro and paste the next 5–10 debug lines if the 1007 still happens? I’ll inspect them and point to the exact field causing Gemini to expect audio.