Universal 2 addresses the part that makes transcripts actually usable: getting names, numbers, and formatting right automatically. The 24% improvement in proper noun recognition and 21% gain on numerical data are not just benchmark wins - they translate directly into transcripts that need significantly less manual cleanup.
What stands out is the formatting layer. Handling punctuation, casing, email addresses, and monetary amounts consistently means the output is ready to work with immediately, not just technically correct. For anyone processing call recordings, meeting transcripts, or interview audio at any volume, that reduction in editing time compounds quickly.
The free tier makes it straightforward to evaluate before scaling, which is the right approach for a tool that lives or dies by how it performs on your specific data.
We also reviewed Universal 2 on our platform https://www.producthunt.com/products/completeaitraining-com — where we list and categorize 7,000+ AI tools for different jobs and skills.
Hey Product Hunt 👋 Harnoor here, with the team at AssemblyAI. Today we're launching AssemblyAI-STT-3.7, a new family of Speech-to-Text models.
Live meetings, voice agents and recordings go wrong in different ways, so this release is three models, each built for one job:
🎙️ AssemblyAI-STT-3.7-Realtime, for live conversation: meeting notetakers, ambient transcription and live captions. Speaker labels arrive in real time, and it code-switches natively across 32 languages. $0.30/hr.
🤖 AssemblyAI-STT-3.7-Realtime-Agent, for voice agents. It's built to stay with the person talking to your agent, ignoring background voices while still catching a quiet "yes". Voice Focus, keyterm prompting and context carryover are built in, across the same 32 languages. $0.45/hr.
📼 AssemblyAI-STT-3.7, for recorded audio. Language coverage grows from 18 to 32, with improved speaker diarization. $0.21/hr.
All three are live in the API and docs now, and you can start free with no credit card.
If you're building a notetaker or a voice agent, what's the one transcription mistake you most want gone? We'll be in the comments all day.
Harnoor and the AssemblyAI team