A transcript can capture every word and still miss what people need: who said what, a translation for a colleague, or the correct spelling of an industry term. On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, a free, open-weight model designed to do all of that in one place, steered by plain-language instructions.
One checkpoint, many speech jobs
Speech pipelines have traditionally stitched together separate tools: one model for recognition, another for speaker diarization (labeling who is speaking), and another for translation. NVIDIA’s pitch is consolidation. According to the company’s Hugging Face announcement, a single checkpoint:
- Transcribes 40 languages.
- Translates 31 of them to or from English.
- Attributes speakers in 13 languages and labels up to eight speakers.
- Handles recordings up to an hour long, with word-level timestamps on request.
- Answers questions about audio (trained in English).
The output is controlled by a prompt. The same meeting recording can become a verbatim transcript, a cleaned-up written version, a speaker-labeled record, or a translation, depending on the instruction. For specialist vocabulary, users can pass likely terms in the prompt before resorting to training.
How it is built
NVIDIA’s research page describes Transcribe as a speech-augmented language model. It couples a Transformer speech encoder and a frozen Nemotron 3 Diarization model with the Nemotron 3.5 Lightning 30B-A3B language model, a hybrid Mamba-Transformer mixture-of-experts design. The result has roughly 4.3 billion active parameters, which is why NVIDIA says it can run on a single chip, from Jetson AGX Thor and DGX Spark hardware up to Blackwell data-center GPUs, with quantized versions for local use.
How it scores
NVIDIA reports that Transcribe ranks first among open-weight models on the English track of Hugging Face’s Open ASR Leaderboard, with an average word error rate (WER) of 5.33%. On multilingual tests it posts a 4.62% mean WER across seven languages, and 10.98% on long-form audio, both the lowest among the open-weight models in its comparison. For speaker attribution, the company says it has the lowest error rate on David AI’s 21-language diarization leaderboard.
TechCrunch notes an important limitation: Transcribe works on recorded audio, not live streams. That makes it a fit for meeting notes, lectures, interviews and archives rather than real-time captioning.
Granary v2 and customization
Alongside the model, NVIDIA released Granary v2, an open speech dataset and curation pipeline, plus fine-tuning recipes in its NeMo Speech tools. The aim is to let developers adapt the model to their own accents and jargon. In a case study using specialized vocabulary, NVIDIA says fine-tuning cut the word error rate on held-out speakers by 45.8% relative to the best prompt-only result.
A practical example comes from Agora, a real-time voice and video platform, which used word-level timestamps to turn a recorded class into a transcript shown next to the replay, so students can see what was said and when.
Why open weights matter here
Because the weights are downloadable, organizations can run Transcribe on their own hardware and choose where their audio is processed. For teams handling sensitive recordings, keeping audio in-house can simplify privacy and compliance, and it avoids per-minute API fees at scale. The release lands in a busy period for voice AI: TechCrunch points to recent launches from Mistral, ElevenLabs and Salesforce.
Practical takeaways
- Start with prompts. Try instructions and term lists for your domain before investing in fine-tuning.
- Spot-check outputs. NVIDIA itself advises checking names, speaker changes and answers against the recording.
- Know the boundary. For live captioning, look at streaming models; Transcribe is built for recorded audio.
- Measure on your data. Leaderboard scores use public test sets; your accents, microphones and vocabulary may differ.
The bottom line
Nemotron 3.5 Transcribe turns a multi-tool speech stack into a single, promptable, open model that can run on one GPU. For anyone building meeting assistants, lecture archives or multilingual note-taking, it lowers both the complexity and the cost of getting from audio to useful text.