AssemblyAI STT and endpointing interaction

Support Help
1

Hi,

We’re experimenting with the AssemblyAI STT provider and, in the interest of accuracy, we’ve set the endpointing setting to 3000ms (max possible).

This very reliably reproduces a double transcription issue. On everything the customer says you get 2 transcripts - first with whatever they said, and a bit after that a second one with a suffix of what they said. In the meantime the agent might have started talking and gets interrupted. This is very easy to reproduce - here are some calls where it happened call_40cb7cc755a20702670e9cd4276, call_6b53241aaf93a97504fa3605869, call_83f6923849fae1a5e2c2ba5b318.

This also happens (but more rarely) with lower endpointing values e.g. with 500ms - e.g. in call_920dd0d19a533a05eb1f943d14f, we got “It’s just a tub.” and then “tub”.

We’ve not observed a similar issue with the other STT providers. Can you please take a look?

2

Hello @plamen Your issue has been escalated to the team for further investigation. We’ll update you as soon as we have more information.

Thank You

3

Hey @plamen The team has pushed a fix for this, and it is expected to be rolled out by the end of the day (EOD).

Thank You

4

Thanks @Shah-Fazal - the duplication issue seems resolved.

However, now we’re hitting the other issue I reported here - AssemblyAI STT high E2E latency even though STT latency is low . We’re getting exteremely high E2E latency even though the individual components are quick e.g. 2026-07-29 07:16:03.412 call_b0682c40c1cbec7b294faa25b73 info: Latency for response_id: 1 - e2e: 4907 asr: 285 llm: 1435 tts: 156

Going to report in the other thread as well. I appreciate you continued help!