GPT-Live starts calls talking about a buffet in a strange voice?

Ive had 2/251 GPT-Live calls (Appliance Repair CSR Ops) that start with a massive artifact where the voice is super exasperated talking about a ‘flexible buffet’, and its weird it happened twice because the script/phrasing IS EXACTLY DUPLICATE on these 2 separate incidents.
Attached are the audio recordings and a matching log file from each call. Please listen to the recordings but you see the transcript doesnt even reflect/pickup this buffet artifact accurately, like it doesnt even know what its saying.
GPT-LiveBug - Google Drive
The logs show my telephony operation. Call comes in, warm up GPT-Live (connection takes a few seconds so we pre-connect it while we play 10 seconds of audio files to let users know about call recording (could be picking up white noise from the audio files but I think the mic is muted, so picking up weird connection static maybe?), we force/encourage a ‘greeting’ line to introduce the AI (shown in transcript) but both of these recording jump straight into talking about "the flexible buffet’ in the middle of a sentence and show the greeting in transcript but never speak it.

It then continues on for a few sentences like its talking to itself with no delay to intake audio. My logs also show when input/output audio is flowing, watch the pattern. The few sentences it runs on with were duplicate in the 2 audio logs. One of the audio logs the user doesnt hangup and asks her what the whole ordeal was about, and she jumps right back into the real prompt/script like it never happened.

Logs show we are sending the prompt fine, shows we are sending the greeting fine. Is this like a debug “Quick brown fox jumps over the lazy dog” AI voice warmup its not supposed to be saying? Cant figure out how it would fail twice and duplicate the same words/sentences/tones. These happened 1 week apart with 0 changes to the underlying code.
“Assistant: {redacted} Appliance, This is Grace.
Assistant: Okay, so maybe we’ll do the Flexible Buffet, and then everyone together for dessert.
Assistant: [gasp] You’re kidding. A five o’clock sit down dinner? Okay, absolutely. That’s fantastic.
Assistant: I just have a couple questions about the buffet”

Is there somewhere else to elevate this too? This seems like a major bug or flaw.

SIP pickup should only take about 2-3 seconds and then your agent should be able to dive in?

I listened to your clip, very funny!

Can you log the audio coming from OpenAI and write it to disk so you can inspect that entirely separately from the telephony? that way you can confirm if it’s coming from OpenAI and being passed on down the telephone.