Live: 3.6 Pro Realtime ranks #1 on Coval and Pipecat
Catch short replies competitors miss. Same endpoint, same price, one line to switch. 45% fewer wrong confirmations on the one-word turns that decide a call.
- Product
- AssemblyAI
- Sent from
- comms.assemblyai.com
- Sent
No dark-mode styles — this email renders the same in a dark client.
Introducing Universal-3.6 Pro Realtime Catch short replies competitors miss. Same endpoint, same price, one line to switch. 45% fewer wrong confirmations on the one-word turns that decide a call. <https://www.assemblyai.com/?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=header> New release <https://www.assemblyai.com/blog/universal-3-6-pro-realtime?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=hero> Today we’re releasing Universal-3.6 Pro Realtime. The accuracy 3.5 Pro delivered on quiet calls now holds on loud rooms, accents, and phone lines. Measured against 3.5 Pro on real agent conversations in production. Pipecat STT benchmark 0.96% Pooled WER, at the front of the accuracy and speed frontier. <https://github.com/pipecat-ai/stt-benchmark?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch> Coval STT benchmark 2.2% WER, and first on their 7-day and 30-day windows. <https://benchmarks.coval.ai/stt?window=7d&utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch> Try it now <https://www.assemblyai.com/dashboard/playground?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=benchmark-card> What you get: Takes no for an answer: 45% fewer wrong confirmations on the one-word turns that decide a call, and 52% fewer in heavy background noise. Background voices stay out: with voice_focus on and a second talker in the room, word error against the caller drops 62%. 32 languages, one endpoint: wrong-language words fall 95%, so most agents can leave detection on automatic. Fourteen languages are new. Turns end when the caller is done: end-of-turn precision is up 3.8 points at the same recall, with median endpoint latency unchanged at 537 ms. Speaker labels improve mid-session: set speaker_labels_revision_interval_ms and revisions arrive during the call, not only at the end. Same price as Universal-3.5 Pro. The change is the model name: set "speech_model": "universal-3-6-pro" on your existing Realtime API connection. Try it now <https://www.assemblyai.com/dashboard/playground?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=price-cta-pair> learn more <https://www.assemblyai.com/blog/universal-3-6-pro-realtime?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=price-cta-pair> <https://www.assemblyai.com/?utm_source=email&utm_medium=owned&utm_campaign=u36-pro-realtime_launch&utm_content=footer> AssemblyAI is the Voice AI infrastructure millions of developers build on. 169 Madison Ave Suite 38365 New York, NY 10016 Unsubscribe <https://links.iterable.com/s/u/HH6lylRS7eofKEI863ADjtxfwLZaIXeKXIGSoAoniCZkQ4QdVoWRbKyfzU7JDJZQFeSXQr63U-42OgCe10FdCLKFAJDtXSgYgABE5Kf3vnJuqMntMF69B7Alj85OPBbIBtP3puA3LegU1a22LSba2oiR5QYAao74fRCT3f2WV3xYalKgS2BmPdEPXpzdrnqzft87GJb-xDLBXll8/9UcA4rlHS7xFtiaN9npVB8EXYqZ8L0As/25>
- Accession
- nse/assemblyai/2026-09-30/2896
- From
- AssemblyAI Team · team@comms.assemblyai.com
- Sent
- 30 September 2026 · 13:00 UTC






