Universal-3.5 Pro - The most accurate STT model from AssemblyAI.

Hey Product Hunt 👋 Happy to be back with another model from the team at AssemblyAI, and today we're launching Universal-3.5 Pro for Async & Realtime.

If you've ever built on top of transcription, you know the transcript is the first mile: everything downstream—summaries, agents, analytics, search—is only as good as the words you start with. So we focused this release on getting that first mile right, especially for the messy, multilingual, real-world audio that most models still stumble on.

Three things we're most excited about:

🌍 Native code-switching across 18 languages. When a speaker moves from English to Spanish and back mid-sentence, Universal-3.5 Pro transcribes the mix in the language it was actually spoken—no separate models, no configuration, no mangled boundaries. It's native to the model across English, Spanish, German, French, Portuguese, Italian, Turkish, Dutch, Swedish, Norwegian, Danish, Finnish, Hindi, Vietnamese, Arabic, Hebrew, Japanese, and Mandarin.

🗣️ Our most accurate speaker diarization yet. Cleaner "who said what," including on short turns, overlapping speech, and noisy environments where diarization usually falls apart. [Drop in the DER / cpWER improvement figure from the post.]

✍️ Contextual prompting. Give the model a plain-language prompt about your audio—the domain, the scenario, the names and jargon that matter—and it biases toward getting those right. No brittle vocabulary lists to maintain.

— Devon & the AssemblyAI team

This is really interesting from a filmmaker/interviewer perspective. I immediately think of long interviews, oral history projects and documentary archives where the real value is not just transcription, but being able to find the exact moment someone said something important.

Curious how well AssemblyAI handles long-form interviews with multiple speakers, accents and imperfect field audio. Do you see makers using this for media archives and documentary workflows too, or is your main focus now voice agents and product teams?

Makes sense it's still cooking. As a stopgap we ended up flagging any segment where the speaker label flipped inside a short window as low-trust, then holding it back from the summary until someone actually glanced at it. Crude, but it cut the silent misattribution bugs a lot. A native per-segment confidence would let us delete that whole heuristic, so glad it's on the radar. Thanks for passing it along to the team.

How does the pricing scale once you start pushing a lot of real-time streaming hours, and are there usage caps or rate limits I should plan around when building a voice agent?

The part that grabs me is transcripts I can actually trust when the audio is a mess. Half the recordings I deal with have people talking over each other and switching languages mid sentence, so this feels genuinely useful, Devon.

The joint diarization and ASR approach resonates. When we piped transcripts into an agent pipeline, our worst bugs weren't word errors, they were speaker mixups: one misattributed turn and every downstream summary that keyed on who-said-what inherited the mistake, silently. Since you produce speaker-change points in the same pass, do you expose a per-segment confidence on the attribution specifically, separate from the transcription confidence? That's the signal I'd want to gate on before letting an agent act on something like 'the customer agreed to X.'

Just tested this with a few recordings, and it’s great. The speaker diarization is very accurate.

Congrats on the launch! The code-switching support is honestly what stood out to me most, built for real conversations instead of clean single-language audio. Also like that the contextual prompting section shows actual before/after examples instead of just claims. Curious how it handles domain-specific jargon at scale, like the medical entity example. Is there a limit to how much context you can feed it before latency takes a hit on the Realtime endpoint?

The PII redaction + LLM gateway combo is the part I would wire in first for anything touching call recordings. Concrete question on the boundary: does redaction happen before the raw transcript is ever persisted or logged on your side, or is it a post-processing pass over a transcript you have already stored? And with contextual prompting to steer results, does the prompt/context I pass get retained or logged anywhere, or is it dropped right after the request?