What if AI could transcribe an entire meeting, distinguish every speaker, translate conversations, and tell you exactly when each word was spoken?
Even when several people are talking over each other?
Meet NVIDIA Nemotron 3.5 Transcribe, a new open-weight speech recognition model designed to handle some of the most difficult problems in audio transcription.
And according to NVIDIA's benchmarks, it's setting new standards for open-weight speech AI.
Here's what Nemotron 3.5 Transcribe can do:
40 languages: Transcribe and process multilingual conversations.
Speaker diarization: Identify who's speaking, even when voices overlap.
Word-level timestamps: Assign precise start and end times to individual words.
Speech translation: Translate spoken content into other languages.
Smart transcription: Remove filler words, clean up grammar and punctuation, or preserve every spoken word.
Audio understanding: Answer questions about the contents of a recording.
Long recordings: Process up to 60 minutes of audio in a single request.
Context-aware recognition: Improve transcription of names, technical terminology, and specialized vocabulary.
And here's where things get interesting.
Most transcription tools struggle when several people speak simultaneously.
Think podcasts, business meetings, interviews, or chaotic group conversations.
NVIDIA designed Nemotron 3.5 Transcribe specifically to handle these situations.
The benchmark results are impressive.
According to NVIDIA's published evaluations:
4.71% word error rate on the English Open ASR Leaderboard.
4.62% word error rate on the multilingual Open ASR Leaderboard.
15.80% speaker-attributed word error rate across several challenging multi-speaker datasets.
These results place Nemotron 3.5 Transcribe ahead of the other open-weight systems NVIDIA compared against.
And there's another interesting technical detail.
Despite its 34 billion total parameters, the model activates only approximately 4.3 billion parameters during inference.
That's thanks to its hybrid Mixture-of-Experts architecture, which combines speech encoding, speaker identification, and language processing.
The model is available in several formats:
BF16 for maximum accuracy
FP8 for reduced memory requirements
NVFP4 for optimized deployment on NVIDIA Blackwell hardware
But there's a catch.
This isn't a tiny speech model you can casually run on any laptop.
The standard BF16 deployment is documented for an NVIDIA H100 with 80 GB of memory, while optimized variants require other high-end GPUs.
And although NVIDIA reports extremely high transcription throughput with its optimized inference stack, those particular performance improvements aren't all included in the public release yet.
Still, this is a significant development for open-weight AI.
Imagine processing hours of interviews, podcasts, meetings, and video footage without manually identifying speakers or synchronizing subtitles.
Or building AI agents that can analyze entire conversations, extract important details, translate speech, and generate searchable transcripts.
For creators, researchers, developers, and businesses working with audio, that's a pretty powerful toolkit.
We're moving from AI that simply converts speech into text to AI that actually understands the structure and context of a conversation.
And increasingly, you can run that technology on infrastructure you control.
Try NVIDIA Nemotron 3.5 Transcribe:
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
Official research and benchmarks:
https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/
#NVIDIA #Nemotron #SpeechRecognition #SpeechToText #OpenWeightAI #AIAudio #GenerativeAI
YouTube / ABVX