Gemini Live API (gemini-3.8-live): dropped sessions and runaway audio
Before sending, fill in: your Google AI Studio / Cloud project number(s) and the contact email.
We run voice agents (tenant phone intake for property managers, English and Spanish) on the Gemini Live API.
On 2026-10-19 we saw three serious failures on gemini-3.8-live: sessions closed by the server mid-conversation that
could not be resumed, replies whose audio kept streaming for over a minute after the words ended, and an output
transcription containing words that were never spoken. A fourth item explains why we could not move to
gemini-3.8-live-extended-thinking.
All times are UTC (US Eastern is UTC−4).
Our setup
- API: Gemini Developer API (Google AI Studio API key), Live API (
BidiGenerateContent), native audio output. - Model:
gemini-3.8-live, temperature 0.7. - Client: LiveKit Agents 1.8.8 with
livekit-plugins-google1.8.8 (Pythongoogle-genai; 2.31.0 in our local
install of the same requirements). - Session settings: tools declared
NON_BLOCKING. Tool responses that need no reply are sent with
scheduling: SILENT. Output audio transcription is on and input audio transcription off. After a dropped
connection the client resumes the session. - Project number(s): ______
1. Sessions closed with 1011, then resumption refused with 1008
Mid-conversation, the server closed the Live session with 1011 Internal error encountered. The client reconnected
0.1 s later and asked to resume the session; the server answered 1008 BidiGenerateContent session not found or expired. The session could not continue and the caller was disconnected. Both times this happened about 1 s after
the client sent content.
| Session started | 1011 received | 1008 on resume | Sent just before |
|---|---|---|---|
| 2026-10-19 18:18:26 | 18:19:13.50 | 18:19:14.06 | a user text turn (18:19:12.35) |
| 2026-10-19 18:19:44 | 18:21:17.40 | 18:21:17.87 | a tool response (18:21:16.30) |
A third session, started at 18:32:57, dropped 20 s in, right after the model’s greeting. We saved its log but have
not reviewed it, so we can’t yet say it was the same error. A session from our other agent on gemini-3.8-live,
18:33:35–18:36:23, did not drop.
Error text, as logged by the client:
error in receive task: 1011 None. Internal error encountered.
Gemini Realtime API error: 1008 None. BidiGenerateContent session not found or expired
AgentSession is closing due to unrecoverable error
Questions: What caused the 1011 errors? Why was the session not resumable seconds later? Was there an incident on
gemini-3.8-live around 18:15–18:35 UTC?
2. Runaway audio: audio keeps streaming long after the words
After the model has said its sentence, the reply’s audio sometimes keeps streaming (near-silent) for a long time.
The caller hears the sentence and then silence. The reply is not over until the stream ends, so nothing else can
happen in the conversation meanwhile.
Clearest example, 2026-10-19 19:22:52–19:24:18: the reply after a tool response was one sentence of about 10 s:
“Your urgent service issue has been submitted. Our team will contact you as soon as possible. If there’s anything
else I can help with, just tell me now. Otherwise I will end the call. Thank you!”
The reply’s usage metrics reported 2,141 output tokens. At 25 audio tokens per second that is about 86 s of
audio, and the reply completed only at 19:24:17.9. Normal replies in the same session: a long read-back was 725
output tokens (about 29 s of audio), and one-line questions were 124–156.
At 17:14:13 a reply with a 2-second goodbye line did not finish for 30 s, when our client stopped waiting and
closed the call. A similar case followed at about 18:35:50.
Questions: Is this a known issue with gemini-3.8-live? Can generation stop when the speech ends? Is there a
setting that limits trailing silence?
3. Output transcription contains words that were never spoken
On a phone call at 2026-10-19 22:38:31, the model’s reply was the question “I’m sorry you’re dealing with this. May
I have your full name?”. The caller heard only that. The reply’s output audio transcription continued with
unrelated text that was never spoken:
“I’m sorry you’re dealing with this. May I have your full name? odors, and other unpleasantries. A clean trash
bin promotes better hygiene, prevents pest infestations, and keeps the surrounding area smelling fresh.”
The transcript is what we store as the call record, so customers see text the agent never said. This was not a
long reply: its usage metrics show 91 output tokens, the same as the next one-line question (“Hi Barb. What’s your
property address and unit number?”, 91 tokens). So the extra words have no audio behind them: the transcription
itself contains text that was not spoken.
Questions: Can the output audio transcription contain text that has no audio? How can we tell such text apart
from spoken words?
4. Extended Thinking cannot be used with silent tool responses
We tried gemini-3.8-live-extended-thinking (thinking level medium). The server closed the session at the first
tool response, at 19:50:24:
1007 None. Function response scheduling is not supported for this model.
Our tools update state (save the caller’s answers) and must not make the model speak again. scheduling: SILENT is
how we prevent that on gemini-3.8-live. Without it, the model may speak again after every tool call, so we cannot use
Extended Thinking.
Request: support SILENT scheduling, or another way to send a tool response that needs no reply, on Extended
Thinking.
Evidence
Client log excerpts with full error lines and timestamps are available on request.