Universal-3.5 Pro: native code switching, our most accurate speaker diarization yet, and expanded language support
Our new flagship async speech-to-text model handles real-world audio the way it actually happens — all at $0.21/hr.
Today we're releasing Universal-3.5 Pro, our new flagship async speech-to-text model. Universal-3.5 Pro handles real-world audio the way it actually happens. Native code-switching across 18 languages captures every word in the language it was spoken. Our most accurate speaker diarization yet ties every word to the right speaker, even through crosstalk and noise. And contextual prompting lets you prime the model with domain knowledge or prior context. All at $0.21/hr.
Real-world audio breaks most transcription. Code-switched conversations default to a single language. Short turns and quick exchanges get misattributed. Background noise corrupts critical details. Universal-3.5 Pro is built for this. It captures code-switched speech natively across 18 languages. It handles complex speaker overlap with better accuracy than anything we've shipped. And contextual prompting means you can guide the model toward your domain.
Native code switching across 18 languages out of the box
Most transcription systems treat code-switching as an edge case. They force everything into one language or mistranscribe the minority language entirely. Conversations split across Indian English and Hindi collapse into one. French in a Canadian call comes back garbled.
Universal-3.5 Pro handles this natively. No configuration needed. No separate pass. The model transcribes every word in the language it was spoken across 18 languages, with stronger accuracy on accented speech than ever before. Because the code-switching is built into the model itself, not stitched on after, the transcript reflects the conversation as it actually happened.
See Universal-3.5 Pro code-switching in action:
English ↔ Mandarin
A linguistics explanation slides between English and Mandarin mid-sentence.
Truth
But this sentence, 我父母不工作了, you can see the 了 at the end of the sentence indicates that the situation now, my parents don’t work, is different from what it was before.
AssemblyAI
But this sentence, 我父母不工作了, you can see the 了 at the end of the sentence indicates that the situation now, my parents don’t work, is different from what it was before.
Competitor
But this sentence, You can see the at the end of the sentence indicates that the situation now, my parents don’t work, is different from what it was before.
See how Universal-3.5 Pro compares on code switching
Normalized word error rate on code-switched audio, across five language pairs. Lower is better.
Code-switching benchmark
Average normalized word error rate
Lower is better
Universal-3.5 Pro
ElevenLabs Scribe v2
Deepgram Nova-3 Multilingual
OpenAI GPT-4o Transcribe
Normalized WER on code-switched audio (lower is better)
The most accurate speaker diarization we've ever shipped
Diarization has traditionally been built as an afterthought. A separate system answers "who is speaking when," working completely independently from the ASR system that answers "what is being said." In the conventional approach, these two streams are stitched together by aligning timestamps. The result is brittle: sentences get split at awkward positions, short turn-taking gets lost, and overlapped speech falls apart.
Universal-3.5 Pro solves both problems jointly. The model produces not just the transcript, but also where in that transcript the speaker changes. The result is a speaker-annotated transcript that follows the natural flow of conversation. It captures short turns and rapid back-and-forth that timestamp-merging approaches miss.
On real call center audio, this matters. Agents don't speak in clean blocks. They interrupt each other, confirm details mid-sentence, and change hands multiple times. Every word assigned to the wrong speaker corrupts the downstream analysis. Universal-3.5 Pro captures this complexity with higher accuracy than anything we've shipped, measured on the metric that actually matters: the combination of transcription accuracy and speaker attribution together.
See Universal-3.5 Pro speaker diarization on real audio:
Emergency dispatch call
Speaker A: Ambulance emergency, which town or suburb?
Speaker B: Redfern.
Speaker A: Okay, tell me exactly what’s happened.
Speaker B: I’m not— just not feeling so good. Chest pain.
Speaker A: How old is the patient?
Speaker B: 51.
Speaker A: Has he ever had a heart attack or angina?
Speaker B: No.
This also changes how diarization should be measured. The field has historically relied on diarization error rate (DER), but DER compares time regions — who was speaking when — and never looks at the words, so it can rank outputs backwards from what any listener would say. In our analysis, a near-perfect transcript with every word on the right speaker scored a 23.4% DER simply because the system didn't label an 11-second group laugh, while an output that misattributed nearly a third of its words and dropped a speaker entirely scored a better-looking 15.1% — those errors are short in seconds, so DER barely notices them. cpWER (concatenated minimum-permutation word error rate) instead measures what users actually see: for each speaker, what fraction of their words did the system get wrong or credit to someone else? We optimized for cpWER, and Universal-3.5 Pro achieves the highest overall speaker diarization accuracy, measured across a wide range of diarization conditions.
How it compares
cpWER across diarization conditions (lower is better)
Contextual prompting
Universal-3.5 Pro is highly accurate out of the box. With contextual prompting, you can push accuracy even higher for the cases that need it. Pass the model a domain or context and it uses that to guide the transcription.
Here's a 2-second clip from a League of Legends pro interview:
League of Legends pro interview
Prompt
This is a League of Legends Pro Interview.
Without prompting
And so look who I’ve been a dear.
With prompting
In solo queue, I ban Azir.
Real-world audio is rarely clean — short clips that offer little context, background noise that buries speech, or domain-specific terms that sound like everyday words without the right framing to tell them apart. Give the model a domain or context and it uses that to guide the transcription.
Clinical: feed in a patient's prior-visit note and the model catches drug and condition names far more reliably; in our internal healthcare test, prior-visit notes cut missed medical terms by 31%, even from an earlier visit.
Meetings: drop in the last meeting's notes or the agenda; topics, project names, and people persist across sessions, so there's no need to hand-extract key terms.
Call centers: product names, brands, competitor names, and plan tiers are known in advance; supply them once as context.
Expanded language support
Universal-3.5 Pro now supports 18 languages at full accuracy, with mid-sentence code-switching so bilingual conversations never pause for the model to catch up. Whether your users code-switch naturally or you need to support multiple languages across your platform, the same model handles it all.
Supported languages: English, Spanish, French, German, Italian, Portuguese, Arabic, Danish, Dutch, Finnish, Hebrew, Hindi, Japanese, Mandarin, Norwegian, Swedish, Turkish, and Vietnamese.
What our customers say about Universal-3.5 Pro
For recruiting platforms, transcription accuracy isn't a feature—it's foundational. When context gets lost in transcription, everything downstream breaks: hiring decisions rely on clean, accurate records of candidate conversations.
Metaview experienced this firsthand. By moving to Universal-3.5 Pro and leveraging contextual prompting to inject meeting metadata (calendar titles, participant names, organizations), they unlocked a measurable jump in transcription quality.
At Metaview, transcription is the first mile of our AI product: if names, roles, companies, and context are wrong, everything downstream gets harder. Since moving to AssemblyAI Universal-3.5 Pro, we've seen a meaningful improvement in the confidence tail of our production transcripts. In a sampled read of our own usage, low-confidence tokens fell by roughly 47%. What stands out is not just the model quality, but the way Universal-3.5 Pro lets us bring real meeting context into transcription, from calendar titles to organizations, domains, and participant names, so recruiting conversations come through with the nuance our customers depend on.
For ambient voice products handling real-time conversations, transcription needs to be rock-solid—both in accuracy and platform dependability. When your infrastructure powers critical workflows, technical excellence is only half the equation; partnerships matter just as much. Commure found this balance with AssemblyAI's suite, building on proven ASR capabilities while exploring new real-time options.
We've integrated the newest models from AssemblyAI for pre-recorded audio ASR in our ambient product, and it's been excellent. We're now exploring Universal-3.5 Pro for async and realtime speech-to-text capabilities for new use cases. What's been just as important is the reliability of the platform itself—both technically and in terms of partnership.
For investment banking and M&A advisory workflows, a single transcription error can cascade into costly mistakes. When deal teams rely on conversation records to track commitments, confirm terms, and document due diligence, transcription precision becomes non-negotiable.
Junior AI powers these critical workflows with conversation intelligence that investment banks, private equity firms, and consulting teams depend on. Accurate transcription of deal discussions—where every figure and commitment matters—is foundational to their product.
Our clients don't have margin for error. A misquoted figure or missed commitment in an M&A process is very costly. AssemblyAI is a key transcription partner, their accuracy is a core reason Junior has become the trusted tool across investment banking, private equity and consulting.
How to access Universal-3.5 Pro
To access Universal-3.5 Pro, simply set "speech_models": ["universal-3-5-pro"] in your API
configuration.
Alternatively, if you want to always automatically upgrade to the latest Universal Pro model, omit
speech_models from your API configuration. If you are currently specifying
universal-3-pro, we'll publish a migration timeline so you can move on your own schedule.
Test universal-3-5-pro in the
Playground
with your own audio, or read the
async guide
to get started.
Universal-3.5 Pro Realtime is also available, and is the new default for realtime transcription and the speech foundation under our Voice Agent API. Test it in the Playground with your own audio, or read the realtime guide to get started.