Hi team, Since yesterday, January 13th, 2026 (afternoon, Spain time), we’ve detected an issue where the transcriptions received in chat completions are incorrect / degraded. We’ve already reviewed our internal logs and infrastructure, and our custom transcriber is working exactly as before (no code changes, same providers, same audio input, same websocket flow). What we’ve observed • The issue starts from the firstMessage of the conversation. • From that point on, the transcription quality is clearly wrong or inconsistent. • This happens even though the audio stream and STT pipeline are unchanged on our side. • Prior to Jan 13 (earlier the same day), everything was working as expected. Our setup (for context) • Custom transcriber running in Cloud Run • WebSocket connection between Vapi and our STT provider • Same language, same audio format, same buffering strategy • No recent deployments or configuration changes on our side Related internal tickets (examples) We’re attaching a few concrete cases where this behavior can be observed: • 019bbba2-21ad-7bb0-be07-ef4e2d119426 • 019bbb9e-dc4b-7aae-af94-b70bfe0ea7e3 • 019bbb71-1d88-7aae-af7a-2bb71377d12d Question Has there been any recent change on Vapi’s side related to: • firstMessage handling • Transcription routing • Chat completions input normalization • Session initialization or buffering logic Happy to provide raw logs, timestamps, or sample payloads if needed. Thanks in advance

@User We are facing the same issue. This is a critical and highly urgent matter, as it is impacting all conversations without exception.

@User same here

Looking into this

Did you solve the problem? Same issue

Thanks for flagging this — our team is actively implementing a code change that will fix the issue affecting custom transcribers as well as all other transcribers. We’re testing the fix now and will share an update shortly.

Hello, any updates on this? Our customers are actively asking for a response and clear timelines regarding this issue.

This should be fixed within the next day or two. Testing phase is complete.

some news?

@giodea are you also using custom transcriber or gladia? please provide a call id for us to check

019bc141-17e9-755e-ad91-2dc93b9b082d

We are seeing the same with italian transcriptions using Cartesia

Could you please let me know whether a refund, credit, or similar accommodation is available? We have had 24 hours of service disruption, yet we are paying the full amount for a service that has not been working. thank you!

You have to manage the transcription and the memory in your backend so that the conversations make sense.

What about Summary / structured output / evals? We have to rebuild everything? No more sense to be in vapi platform in this case

Do you know other providers?

a lot: 11labs / retell / livekit / openaiagents and more

Thanks, but 11labs is multi provider? I don't know

Yes for llms, not for tts / stt

I need all: llms, tts and stt. ¿Do you know anyone?

We apologize for the wait. While our engineers work on implementing this fix into our codebase, we recommend switching to deepgram as the transcriber temporarily. We will also credit your account with a one-time coupon.

We have the same problem since the last days. Is there any way to recover the transcript from you logs or do we have to send all audios to the transcriber service again?

Any updates?

please, any update on this?

The fix has been implemented into the daily updates channel. If you are still experiencing issues with custom transcribers, please let us know and provide a call id for us to review.