STT transcript delay/stitch + TTS stutter in eu-central

[LaVivien]

Bug Description

I've been seeing intermittent glitches in our voice AI app since Sep 13, 2026. It is hosted in eu-central (LiveKit Cloud hosted agent)

The major symptoms are :

1. STT transcript display: sometimes words cannot show up one by one smoothly. Instead they "flash", then 1-2s later several words/sentences stitch together and appear as one block in the transcript UI. Sometimes 2-3 sentences merge into a single transcript entry.
2. TTS: occasionally stalls mid-sentence for 1-2s, then resumes, this might be caused by the delay of incoming STT. The TTS audio also sometimes speeds up briefly, then back to normal.

I have spent a day isolating the cause.

• No code or dependency changes on our side around Sep 13.
• I have swapped different TTS and STT models, the symptoms are persistent.
• Ran on different devices, in different frontend hosting env. The same symptom reproduced.
• Ran the same agent code on a laptop, same symptoms reproduced.

All these point to something in LiveKit's agent, maybe text-stream layer. I wonder if there is known latency or delivery issues, or any recent changes that affect STT transcription event delivery timing.

SPECIFIC QUESTION — EU-CENTRAL
Our agent is deployed in eu-central. LiveKit had an incident in eu-central on Sep 5, 2026 (agent join latency, resolved same day). Our symptoms started Sep 13, a week later, we wanted to ask: is there any chance our project's eu-central agent instances have been affected by a related or follow-on issue since Sep 5.

Expected Behavior

The STT transcript should display word by word in the order of people speak.

Reproduction Steps

1. Open https://practice.parlez.app/. login with a gmail account.
2. click any card, and speak in French. the green transcript display what you said .
3. Speak for a few minutes, you will notice the words in green transcript flashes. Wait for a second to two, they stitch together as a whole sentence and display at once. It seems the words didn't arrive in the right order, then finally stitch together.

Operating System

iOS, Samsung, windows11

Models Used

STT: Speechmatics (language=fr, operating_point=STANDARD) - TTS: Google Cloud (WaveNet) - LLM: openAI "gpt-4o-mini"

Package Versions

livekit-agents                           1.5.8
livekit-api                              1.0.7
ivekit-plugins-google                   1.5.8
livekit-plugins-speechmatics             1.5.8
livekit-plugins-turn-detector            1.5.8

Session/Room/Call IDs

session ID, RM_YdQL9PFTtbaW

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response