What's new in Universal-3 Pro: smarter code-switching, faster turnaround, and better timestamps

Universal-3 Pro is the most accurate model on the market, and recent releases push the lead even further. 

We've made meaningful improvements to Universal-3 Pro’s code switching, disfluencies, turnaround time, diarization, and timestamps. Everything below is live. If you’re already using Universal-3 Pro, no code changes are required.

Code-switching: better multilingual transcription out of the box

Mixed-language audio is one of the hardest problems in speech-to-text. Universal-3 Pro now handles code switching significantly better, with consistent gains across multilingual benchmarks.

Spanglish Audio:

Play 1x

0:00

0:00

‍

The result: a ~19% relative WER improvement on code-switching benchmarks, gains across all 20 CommonVoice and FLEURS test sets, and notable improvements on Spanglish audio (Miami corpus). Accented English and medical terminology accuracy improved alongside the multilingual gains. Customers using a custom prompt parameter are unaffected and can continue tuning to their workload.

If multilingual or accented audio is part of your workload, Universal-3 Pro is non-negotiable.

Disfluencies: now supported on Universal-3 Pro

Capturing filler words, hesitations, repetitions, and false starts is critical for verbatim use cases like voice agent training data, therapy session documentation, and compliance recordings. Universal-3 Pro now supports disfluencies through a single parameter.

Audio with Disfluencies:

Play 1x

0:00

0:00

Well, I mean, they may have exaggerated a little bit, but— probably, um, but, uh, no, I think, I think he actually just kind of looked like that in reality.

The result: ~5.9% WER improvement on verbatim datasets 

Accurately capture filler words and repetitions while maintaining transcription quality. For verbatim workloads, Universal-3 Pro delivers what you need in one API call.

Turnaround time: the fastest model in the lineup

Universal-3 Pro now delivers the fastest turnaround time of any model in the AssemblyAI lineup. Recent infrastructure and pipeline work has driven meaningful gains across audio durations:

• P50 latency improved by up to 30%
• P99 latency improved by up to 34%

For long-form transcription workloads, batch processing pipelines, and any use case where turnaround time matters, Universal-3 Pro is now the right default.

Diarization: meaningful accuracy improvements

Speaker diarization is one of the most requested areas of improvement from customers, and we've shipped a wave of upgrades to Universal-3 Pro:

• Short files (under 2 minutes): 19% relative improvement in correct speaker count and 6% relative cpWER improvement.
• Sentence-level speaker consistency: prevents the last word of a sentence from being mis-assigned to the next speaker. 
• Better handling of short segments: improved accuracy on backchannels and single-word responses like "Yeah" or "Okay."
• More accurate speaker assignment: short-utterance assignment is more consistent across multi-speaker audio.
• Max speakers raised to 30 for long files: Increased max speakers detected to 30 speakers for files over 10 minutes.

Timestamps: significantly more accurate, especially for non-English

We've shipped several improvements to how Universal-3 Pro calculates timestamps, balancing accuracy gains against turnaround time impacts:

• English timestamp precision: +15.3% at the median, +15.0% at P99
• Non-English timestamp precision: +8.6% at the median, +58.4% at P99

For captioning, search, alignment, and any downstream workflow that depends on word-level timing, Universal-3 Pro is now the strongest option, particularly for multilingual content.

What's next

These are some of the most meaningful accuracy and quality-of-life improvements we've shipped to Universal-3 Pro since launch, and we're continuing to push across all fronts.

If you're already using Universal-3 Pro, you're getting these improvements automatically. If you're still building on Universal-2, Universal-3 Pro is now the better choice across accuracy, turnaround time, diarization, and timestamp precision. 

If you haven't tried it yet, sign up for a free API key and start building. You get $50 in free credits, no credit card required.

‍