Skip to main content

NVIDIA's New Speech Model Handles 8 Voices at Once

NVIDIA's New Speech Model Handles 8 Voices at Once

NVIDIA has quietly dropped a new open-source speech transcription model, and it's a beast. Called Nemotron 3.5 Transcribe, it's built for professional use—think meeting minutes, interviews, and multilingual content—and it's already turning heads with its ability to separate multiple speakers and handle long recordings.

Under the Hood

The model uses a Mixture-of-Experts (MoE) architecture with a whopping 34 billion total parameters, but only 4.3 billion are activated per forward pass. That means it's both powerful and efficient—no need for a supercomputer to run it. And it can chew through up to one hour of continuous audio in a single pass, so you can throw that marathon board meeting at it without chopping it into pieces.

Speaking in Tongues

Nemotron 3.5 Transcribe supports 40 languages for transcription and can translate between 31 of them and English. That's a big deal for global teams and content creators. But the real showstopper? It can distinguish up to 8 different speakers simultaneously—even when they're talking over each other. No more "Speaker 1" chaos in your transcripts.

By the Numbers

On the Open ASR English benchmark, it posts a word error rate (WER) of just 5.33%, the best among open-weight models. And in Artificial Analysis's non-streaming test, it drops to 2.8%, tying for first place on the open-source leaderboard. Those are numbers that make transcription nerds sit up and take notice.

So if you're tired of garbled meeting notes or spending hours untangling multilingual recordings, NVIDIA's latest offering might just be your new best friend. It's open-source, so you can take it for a spin yourself.

Key Points:

  • NVIDIA open-sources Nemotron 3.5 Transcribe, a speech model with 34B parameters (4.3B active).
  • Handles 1-hour audio in one pass and supports 40 languages, with translation to/from English for 31.
  • Can separate up to 8 speakers in real time, even with overlapping speech.
  • Achieves 5.33% WER on Open ASR English and 2.8% on Artificial Analysis, leading open-source models.