Skip to main content

Kaer NewsSoftware

Nvidia releases Nemotron 3.5 Transcribe model for 40 languages

Nvidia released Nemotron 3.5 Transcribe, a free open-weight model that transcribes 40 languages and runs on a single chip, alongside the Granary v2 dataset, on 28 October 2026.

3 min read
Nvidia releases Nemotron 3.5 Transcribe model for 40 languagesPhoto: TechCrunch

Key points

Nvidia released Nemotron 3.5 Transcribe, a free open-weight model that transcribes 40 languages, translates 31, and labels up to eight speakers.

Nvidia released Nemotron 3.5 Transcribe on 28 October 2026, a free, open-weight speech-to-text model that transcribes audio in up to 40 languages. The company also released version 2 of Granary, its open-source speech dataset. The model is available on Hugging Face and ModelScope, and it handles recordings up to an hour long while telling up to eight speakers apart.

The release matters because it packs transcription, speaker attribution, translation, and audio question answering into a single checkpoint that developers can run without paying licence fees. Nvidia says developers can combine the model, the updated Granary dataset, and its NeMo Speech tools to teach an application an industry's jargon in each language. That targets developers building business meeting and medical transcription tools, two of the most established uses of artificial intelligence in the workplace.

What the model can do

Transcribe translates 31 languages to and from English and attributes speakers in 13. Developers can steer its output with plain-language instructions before it listens, asking it to label speakers, translate a recording, or answer questions about what decisions a meeting produced. In a healthcare example published by Nvidia, a prompt can supply terms such as HbA1c and metformin, telling the model to use them only when the audio supports them.

The model's design explains how it stays small enough to run on one chip. Transcribe pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, so only about 4 billion of its 30 billion parameters are active per token. Nvidia says it can run on a single chip, from the Jetson AGX Thor robotics controller and DGX Spark to Blackwell datacenter chips.

On Hugging Face's Open ASR Leaderboard, which measures transcription accuracy as word error rate (WER, the share of words a model gets wrong) on public test sets, Transcribe ranks first among open-weight models in English at 5.33% average WER, with application programming interface models excluded. Its mean WER is 4.62% across seven languages in the multilingual track and 10.98% in the long-form track, both the lowest in the model card's comparison.

Benchmark rankings and word error rates

Artificial Analysis, which ranks speech-to-text models independently by its own word error rate index, has Transcribe tying for first among open-weight models at 2.8% AA-WER, using about 4 billion active parameters against 24 billion for the model it ties. On AA-AgentTalk, a held-out voice-agent speech set making up 50% of that index, Transcribe scores 1.8% WER versus 2.3% for the next-best open-weight model.

For speaker attribution, Transcribe posts the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison. Diarization means sorting which person said which words in a recording. Nvidia's own medical case study found that fine-tuning cut word error rate on held-out speakers by 45.8% compared with the best prompt alone.

Transcribe works with recorded audio, not livestreams, leaving real-time voice work to rivals. In the same period, Mistral shipped Voxtral 3, ElevenLabs launched Realtime for instant speech in voice agents, and Salesforce released Launa-Omni. Nvidia's own smaller open-weight voice model, Parakeet, has 1.1 billion parameters. Quantized versions of Transcribe, which run models with reduced precision to save memory, are coming soon.

What the model still cannot do

Agora, whose real-time voice and video platform powers live conversations in apps, used Transcribe's word-level timestamps in a classroom demo, turning a recorded class into a transcript shown alongside the replay. Nvidia's model card recommends presets, a full prompt guide, and a task-by-language matrix, noting that audio question answering is trained in English, and advises spot-checking names, speaker changes, and answers against the recording.

Frequently asked questions

What is Nvidia's Nemotron 3.5 Transcribe?

Nemotron 3.5 Transcribe is a free, open-weight model released by Nvidia on 28 October 2026 that turns recorded speech into text. It transcribes 40 languages, translates 31 to or from English, labels up to eight speakers, and handles recordings up to an hour long.

Can Nemotron 3.5 Transcribe handle live audio?

No. Transcribe works with recorded audio, not livestreams, leaving real-time voice work to rivals such as ElevenLabs, which launched Realtime for instant speech in voice agents.

More from this edition