Hey All,
We heard your feedback and this gemini-2.5-flash-native-audio-preview-09-2025 model is much better at function calling and has reduced audio cutoffs. Can you please try again with the new model and let us know what you think?
Thank you!
Thanks for the update - I gave it a try and it does seem to be more usable than before, notably no freezing in the middle of a conversation and functions are called. Sometimes it does seem to freeze on startup. The biggest issue right now seems to be the audio has quite a bit of stuttering or echoing - the most obvious example is when it literally repeats a sentence, though I think the audio “blips” are just smaller forms of that sort of repeated audio. While there could be an issue in my websocket audio pipeline, I have completely crisp audio with gemini-live-2.5-flash-preview so I feel that shouldn’t be the case. While the better pronunciation of native audio is nice, because the general audio quality of gemini-live-2.5-flash-preview is so much higher it seems to work better for me. This is besides the fact that native audio also has been much less reliable at following system instructions, whereas gemini-live-2.5-flash-preview is very reliable in behavior. This may be improvable with prompt engineering but I wouldn’t try it before the audio quality improved.
We noticed the same. The models is still being cut off, now it tries to recover, but it is still to much from being used for any kind of customer communication
As the comment above me said, the model is cutting off, some times trying to recover by speaking the same sentence. This is far from the ideal.
También encontré el mismo problema, pero lamentablemente no encontré una solución.
We are still experiencing the same problem. Are there any updates?
Hey All,
Thank you for taking the time to test the model. We appreciate your efforts in helping us identify potential issues.
To effectively investigate and debug the problem you encountered, it would be extremely helpful if you could provide specific examples: the audio file, the incorrect transcript and the correct transcript. This comparison is invaluable for improving accuracy.
I think we’ll be hard pressed to provide such data since in the end this is Gemini - not providing proper support through Vertex AI is the first fatal mistake Google made.
Now I got this, saying change models from the half cascade model by December.
We have recently launched a new, updated preview version to replace the previuos ones: Gemini 2.5 Flash Native Audio Preview (September 2025 version) (
gemini-2.5-flash-native-audio-preview-09-2025)
As you know, this model is not usable so this is a last straw. Are you developing models in a complete research oriented Deepmind black box or trying to create actually usable models? The former I guess we know. Personally I have no idea how you will get any good faith back at this point.
I’m having the same problem. I’m getting terrible results from the new gemini-2.5-flash-native-audio-preview-09-2025 in general: breaks up mid-sentence, gets stuck, speaks nonsense and so on. I’m worried that they’re considering this generally available and deprecating the other models on Dec 9.
I’ve been using gemini-2.5-flash-native-audio-dialog in my application for quite a while. Previously, there were only 5 sessions per day, but the limit was recently increased. Yesterday, it was disabled. It had some issues with tool initialization — when switching languages, the API stopped sending requests to the tools, and I had to reinitialize them manually. Sessions sometimes got interrupted, and occasionally the model would repeat itself.
Still, the model is excellent and works perfectly for presentations and consultations. It’s available within my quotas. It’s a great example of what neural networks can do today.
Of course, I could build a regular STT–TTT–TTS chain (I’ve already implemented that in Gemini Radio), but I fell in love with this particular model the moment it was released. Competitors don’t have anything like it, which is why it’s so important that this model be deployed — for example, in clinics for entry consultations, at train stations, and in supermarkets. I’d like to implement something like that here in Ukraine.
I don’t have a powerful GPU to train my own model, but Google is very generous to its users, and I can do it in the cloud.
unfortunately it still happens even with gemini-2.5-flash-native-audio-preview-12-2025
Ive tried a system prompt in ai studio:
Act as a breathing coach and guide me through a 4-7-8 breathing exercise for exactly 2 minutes.
Here is the structure:
Inhale through the nose for 4 seconds.
Hold the breath for 7 seconds.
Exhale forcefully through the mouth for 8 seconds.
Instructions for you:
One full cycle takes 19 seconds. Please conduct 6 full cycles to cover approximately 2 minutes.
For each phase, explicitly write out the action and count the seconds line-by-line so I can follow the rhythm (e.g., ‘Inhale: 1, 2, 3, 4’).
Start immediately with ‘Get ready…’ and then begin the first cycle.
Don’t say “cycle”
Critical rule: your tone of voice and pace should be calm and relaxing like whispering or carefully breathing!!
it gives transcript, but audio always missing last few words
here an audio with fun example of how audio works for counting: 
play audio
it doesn’t pronounce number “4” few times !!
Trancript:
Get ready… Inhale: 1, 2, 3, 4 Hold: 1, 2, 3, 4, 5, 6, 7 Exhale: 1, 2, 3, 4, 5, 6, 7, Inhale: 1, 2, 3, 4 Hold: 1, 2, 3, 4, 5, 6, 7 Exhale: 1, 2, 3, 4, 5, 6,
I’ve tried the latest model (gemini-2.5-flash-native-audio-preview-12-2025) and the problem is still there. It seems the quality of the live API’s preview version is much worse than other preview versions. It hasn’t been fixed for months, 
and the crucial part is that they also shut down the model that was previously working.
We are facing this too. It started happening recently (past week) on the Live API. The audio generation is unreliable and cuts off constantly while the text response remains correct. This makes the Native Dialog feature very difficult to rely on for production traffic right now. Hope to see a fix soon.
Hi all,
I’m completely new to Google’s Gemini ecosystem and started using Gemini 2.5 Native Audio with the Live API about a week ago. I’m working in Python (VS Code), and everything was functioning well until early this morning, when it suddenly stopped working.
I’m now consistently getting a “cannot generate a response” error. I didn’t intentionally change anything in my code right before this started happening. The Live API connection still initializes, but response generation fails.
I’ve tried reviewing tracebacks and troubleshooting on my own (including getting help understanding the Python errors), but I’m still a beginner with Python and I’m not sure whether:
-
the model was deprecated or disabled,
-
there was a breaking change,
-
or I should be switching to a different model.
Questions:
-
Is
gemini-2.5-flash-native-audio-preview-12-2025still supported? -
Which model is currently recommended for Live API + native audio?
-
Is there a migration path or known issue related to this?
Any guidance would be greatly appreciated. Thanks!
I’ve been using gemini-2.5-flash-native-audio-dialog in my application for quite a while. Previously, there were only 5 sessions per day, but the limit was recently increased. Yesterday, it was disabled. It had some issues with tool initialization — when switching languages, the API stopped sending requests to the tools, and I had to reinitialize them manually. Sessions sometimes got interrupted, and occasionally the model would repeat itself.
Still, the model is excellent and works perfectly for presentations and consultations. It’s available within my quotas. It’s a great example of what neural networks can do today.
Of course, I could build a regular STT–TTT–TTS chain (I’ve already implemented that in Gemini Radio), but I fell in love with this particular model the moment it was released. Competitors don’t have anything like it, which is why it’s so important that this model be deployed — for example, in clinics for entry consultations, at train stations, and in supermarkets. I’d like to implement something like that here in Ukraine.
I don’t have a powerful GPU to train my own model, but Google is very generous to its users, and I can do it in the cloud.