Nemotron 3.5 Transcribe: 40 Languages, One GPU, Transcription You Control
Nvidia dropped Nemotron 3.5 Transcribe on October 28—a 30-billion-parameter speech model that handles transcription, translation, and speaker labeling all at once, runs on a single GPU, and doesn't phone home. It's open-weight under the OpenMDW-1.1 license and already tops Hugging Face's English speech-recognition leaderboard. For creators and developers who need transcription you control, this changes things.
What It Actually Does
Instead of stitching together separate tools for transcription, diarization, and translation, Nemotron takes a prompt: "Transcribe this and label each speaker," or "Translate to English," or "Give me a clean transcript without filler words." The same model handles all of it. It transcribes 40 languages, translates 31 to or from English, identifies up to eight speakers in 13 languages, and can answer questions about recordings. Hour-long files. One checkpoint.
The catch you need to know: it only works on recorded audio, not live streams. If you're transcribing podcasts, meetings, or video shoots after the fact, you're good. If you need real-time captions for a livestream, this isn't the tool.
What People Are Building With It
Nvidia's own examples focus on practical workflows. Agora, a live video platform, used the word-level timestamps to build a classroom replay tool: your teacher's recording sits beside a transcript, and students can click any word to jump to that moment in the video. That's searchable, accessible teaching.
The model is built for medical dictation and meeting tools—industries where transcription accuracy matters and keeping audio in-house matters more. A developer can now fine-tune it on domain-specific vocabulary (medical terms, legal jargon, product names) in about 14 GPU-hours on a single Nvidia GPU.
For video creators and editors, this opens the door to local captioning workflows. Instead of uploading your footage to a cloud service, paying per minute, and worrying about privacy, you run the model on your own hardware. Multi-speaker workflows get cleaner because the speaker labels are built in, not bolted on afterward.
The Numbers
Nemotron 3.5 Transcribe is the best open-weights model on the English and multilingual Open ASR Leaderboards, with a mean word error rate of 4.71% on the 9 public English test sets. It claims to beat rivals on benchmarks, including Hugging Face's Open ASR Leaderboard. For context, Whisper-large-v3 scores 6.66% WER—so Nemotron wins.
The model has 34B parameters, but only about 4B are used for each step, which keeps it fast and cheap to run. That architecture choice (a Mamba-Transformer mixture-of-experts hybrid) means it fits on a single high-end GPU, and quantized variants (FP8, NVFP4) fit smaller hardware.
What's Actually Good
One open model replacing a pipeline of separate tools is genuinely useful. It works on a single Nvidia chip, from small Jetson boards to data-centre GPUs, and its licence allows commercial use. You can ship products with this.
The speaker attribution (called diarization in the industry) is the release's strongest point. Nvidia models hold the top two spots on David AI's multilingual diarization leaderboard: Nemotron 3.5 Transcribe posts the lowest overall diarization error rate (DER) at 18.26% across 21 languages. That means "Speaker 1" and "Speaker 2" labels are accurate, even when people talk over each other.
The promptable interface is a nice touch. Instead of fixed settings, you tell it what you want in plain language. Examples include "Transcribe this recording and label each speaker", "Translate this audio into English", a list of likely domain terms such as drug names, and "Transcribe without disfluencies" for a clean rather than verbatim transcript.
What Doesn't Work Yet
Transcribe works with recorded audio, not livestreams. That's a real limitation if you're building real-time caption tools. Its performance on Mandarin far-field meetings (34.95% character error rate) is weak, so if you're transcribing Chinese-language meetings recorded from a distance, expect more errors.
Audio Q&A was only trained in English, even though the model speaks 40 languages. Nvidia's own model card warns that language support differs by task.
How to Try It
Agora's blog walks through how they built a classroom replay with word-level timestamps, if you want a worked example. A developer shared a repository and notebook that walk through a local audio clip, seven prompt patterns, the language breakdown, and the raw JSON payloads.
Transcribe is available as a Hugging Face demo, where developers can download a checkpoint and adapt it for domain expertise with fine-tuning. Developers can serve the model with vLLM or run it locally with NeMo-Speech.cpp. The weights are on Hugging Face.
Granary v2, Nvidia's open speech dataset, shipped alongside it: 463,062 hours of speech in 48 languages, which you can use to fine-tune the model on your own vocabulary.