And no minimums: you're billed for the audio you send. Short utterances don't get rounded up to a billing floor, which matters when your workload is thousands of 5-second commands.
It's also a natural fit for voice agents. When an agent needs a complete utterance transcribed between turns, like a captured voicemail, a confirmation, or a quick voice input mid-flow, Sync returns the full text in one shot, no stream to manage.
Sync is the tier where dictation lives. Voice notes, clinical documentation, voice commands, form-fill: anywhere someone talks in short bursts and expects the words to just be there.
➡️ Realtime: text appears while they speak (live captions, real-time conversation)
➡️ Sync: text arrives right after they stop (~134ms)
➡️ Async: text is ready whenever the job's done (batch, long files)
The Sync API is here.
Synchronous speech-to-text, built for the moment right after someone stops speaking, with text back in as few as ~134ms.
Speech-to-text now comes in three speeds:
Speech-to-text is more than a transcription layer—it's a key part of how voice agents understand users.
This article by @AssemblyAI explores why speech recognition quality has a major impact on voice agent accuracy and user experience:
We mispronounced every single item on the menu. Universal-3.5 Pro Realtime got all of them right anyway.
👇 Check out @martschweiger tutorial where he spun up a restaurant ordering agent on @livekit to put our latest model to the test.
He ordered escargots à la bourguignonne,
Our team can't stop talking about the World Cup—so @martschweiger pointed our newest model, Universal-3.5 Pro, at a bilingual Cristiano Ronaldo interview to see how it handles the parts most models fumble.
He switches between Portuguese and English mid-sentence. Universal-3.5
Universal-3.5 Pro Async is here — native code-switching across 18 languages, our best speaker diarization yet, and contextual prompting. All at $0.21/hr.
Learn more:
Ordering a prescription by voice leaves no room for "close enough." Drug names, dosages, patient details—every word has to land.
@stefanjblos built a voice agent that does exactly that—and gets the messy, high-stakes language of medicine.
💊Building agents that work with medical data is incredibly hard. There's a lot of very specific terms and names + very sensitive data + a high-risk environment.
Our resident DevRel @stefanjblos build a prescription ordering agent that is incredible at understanding and
Clicky by @FarzaTV is powered by @AssemblyAI Universal 3 Pro streaming model. I cloned it and upgraded it to Universal 3.5 to show how our newest realtime model can actively listen with context.
The page itself becomes context for the speech model.
As you scroll, Clicky updates
