NVIDIA Nemotron 3.5 Transcribe
Most speech-to-text tools do one job. Nemotron 3.5 Transcribe does several in one model: it transcribes recordings in 40 languages, labels who said what, adds word-level timestamps, translates to or from English, identifies the language, and answers questions about the audio. NVIDIA released it on October 28, 2026 as open weights on Hugging Face and ModelScope, under the OpenMDW-1.1 license, which allows commercial use.
It is a model, not an app. There is no meeting recorder or editor attached; developers build those around it, or call it through a hosted provider. And it is built for recorded audio. NVIDIA says plainly that it is not a live-streaming model.
What One Checkpoint Covers
Older transcription setups often chain separate systems: one model to transcribe, another to work out who spoke when (diarization), a third to translate. Nemotron 3.5 Transcribe handles these in one pass, and you steer it with a prompt. According to NVIDIA, it can:
- transcribe speech, including recordings where speakers switch languages mid-sentence
- attribute speech to up to eight speakers in a recording
- add word-level timestamps
- translate speech to or from English
- identify which language is being spoken
- answer questions about a recording, in English
- use context you supply, such as names and product terms, to spell things correctly
- tag or redact personal information in the transcript
Because it takes instructions, you can ask for more than a plain transcript: “transcribe this, label the speakers, and keep the product names exactly as written in this list.”
Specs
| Nemotron 3.5 Transcribe | |
|---|---|
| Developer | NVIDIA |
| Released | October 28, 2026 |
| Type | Offline speech model: audio in, text out |
| Architecture | Hybrid Mamba-Transformer mixture-of-experts language model with an audio encoder and a separate speaker encoder |
| Parameters | About 34B total including the audio encoder; about 4.3B active per token |
| Context window | 65,536 tokens |
| Recording length | Up to about one hour |
| Input | 16 kHz mono audio in WAV, FLAC, MP3, or OGG |
| Output | Text, with optional word timestamps and speaker tags |
| Languages | Transcription in 40; translation to or from English in 31; speaker attribution in 13 |
| Precision variants | BF16, FP8, NVFP4 |
| License | OpenMDW-1.1 |
About the parameter count. The model’s name says 30B-A4B, and NVIDIA’s announcement describes about 4 billion active parameters out of 30 billion. The model card counts about 34 billion total and 4.3 billion active once the audio encoder is included. Use the model card’s figures for hardware planning. The difference between total and active matters too: active parameters set how much compute each token costs, but the total sets how much memory you need, because the whole model has to be loaded.
Languages
NVIDIA groups the transcription languages into tiers by quality. The top tier is Czech, Dutch, English, French, German, Hindi, Italian, Korean, Polish, Portuguese, Russian, Slovak, and Spanish. A second “production-ready” tier includes Japanese, Swedish, Danish, Finnish, Turkish, Ukrainian, Thai, and several Central and Eastern European languages. A third “supported” tier includes Arabic, Mandarin Chinese, Hebrew, Indonesian, Vietnamese, Norwegian, and Bengali, Marathi, Tamil, and Telugu.
Support differs by task, not just by language: speaker attribution covers 13 languages, and audio question answering and instruction following were evaluated only in English. Check the language matrix on the model card before you depend on a specific task in a specific language.
Nemotron 3.5 Transcribe and Granary v2
NVIDIA announced the model alongside Granary v2, and the two are easy to mix up. Granary v2 is a dataset and a data pipeline, not a model. It contains about 463,000 hours of speech across 48 languages from 16 source corpora (the first Granary covered 25 languages). For each recording it provides transcripts in spoken and written form plus a cleaned version, along with captions, question-and-answer pairs about the audio, context terms, and translations with quality scores. The pipeline that builds it runs on NVIDIA NeMo Curator.
The relationship is simple: Granary v2 is one of the main sources of training data behind Nemotron 3.5 Transcribe, alongside public corpora, licensed third-party data, and synthetic audio. Teams can also use it to fine-tune the model for their own domain. NVIDIA’s announcement does not state a license for the Granary manifests, and the source audio has to be downloaded separately under each original corpus’s license, so check those terms before training on it.
Running It
NVIDIA’s recommended setup is a single H100 80GB GPU running vLLM with NeMo Speech. The model card also lists Hugging Face Transformers, SGLang-Omni, and GGML. GGML builds are aimed at an RTX 5090 or RTX PRO 6000 and are the only way to run it natively on Windows; otherwise Linux is the main platform, with Windows through WSL2. Hardware on NVIDIA’s compatibility list ranges from data-center GPUs (A100 80GB, H100, H200, B200, GB200, GB300) to DGX Spark and Jetson Thor, with the L40S 48GB listed for the FP8 version only.
Open weights do not make it a laptop app. With about 34 billion parameters to hold in memory, the full-precision model needs a data-center-class GPU, and even the compressed builds target the most expensive workstation cards. The quantized FP8 and NVFP4 versions reduce memory use. NVIDIA reports that a forthcoming NVFP4 build processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell GPU, measured on 30-second clips at high concurrency. That is NVIDIA’s own test, under batch conditions that favor throughput.
Not every runtime supports every feature. SGLang-Omni does not output timestamps, speaker labels, or translations, and the optional DFlash2 speed-up turns off word timestamps. A long recording can produce around 20,000 output tokens, so you may need to raise the output limit.
Hosted options: Baseten, DeepInfra, and Simplismart offered hosted inference at launch, each at its own price, and NVIDIA lists the model on build.nvidia.com. A hosted provider avoids the hardware, but your audio then goes to that provider, so check its data terms.
Fine-tuning: NVIDIA publishes NeMo Speech tutorials, and ModelScope supports fine-tuning through ms-swift. NVIDIA’s own medical case study shows why it can matter: on a medical speech test set from Eka Care, word error rate was 30.8% out of the box, 23.8% with a better prompt, and 11.8% after fine-tuning on about 245 hours of accented, medical, and synthetic speech, using roughly 14 GPU-hours. Those are vendor-reported numbers, but the pattern is common: general models struggle with specialist vocabulary until they are adapted. See fine-tuning and LoRA for the techniques.
The License
OpenMDW-1.1 is a permissive license from the Linux Foundation, written for model weights and the materials around them. In practice:
- you can use, modify, and redistribute the model, and NVIDIA’s model card says it is ready for commercial or non-commercial use
- if you distribute it, you must include a copy of the license and keep the copyright and origin notices
- the license places no restrictions on the outputs, such as the transcripts you generate
- your rights end if you sue claiming the model infringes a patent or copyright, unless you are responding to a suit brought against you first
- the model comes “as is,” and the license makes you responsible for clearing third-party rights and obtaining any necessary consents
Poolside uses the same license for Laguna S 2.1. For the general difference between downloadable weights and full open source, see open weights.
How Good Is It?
Speech recognition is usually scored by word error rate (WER): the share of words the model gets wrong, so lower is better. NVIDIA’s headline results:
- Hugging Face Open ASR Leaderboard (English): first among open-weight models at launch, at 5.33% average WER, according to NVIDIA.
- Multilingual and long-form tests: a 4.62% mean WER across seven languages and 10.98% on long recordings, the lowest among the open-weight models in NVIDIA’s own comparison table.
- Third-party evaluations NVIDIA cites: a tie for first among open-weight models on Artificial Analysis’s speech-to-text benchmark, and first among open-weight models (second overall) on David AI’s combined transcription and speaker-attribution benchmark.
Treat these carefully. NVIDIA chose which results to publish and which models to compare against. The outside evaluators are separate organizations, but the model card also lists data licensed from David AI among the training sources, which is worth knowing when reading that particular benchmark. Leaderboards use public test sets such as audiobooks, parliamentary speeches, and earnings calls. Your meeting recordings, with crosstalk, accents, background noise, and internal jargon, may score very differently, as the 30.8% medical result above shows. Strong leaderboard placing makes it worth testing, not a safe default.
Practical Uses
The model fits work where recordings are processed after the fact:
- Meeting and interview transcripts with speaker labels and timestamps, for teams that want to keep audio on their own servers instead of sending it to a meeting-notes service
- Multilingual recordings, including speakers who switch languages, plus translation to English for review
- Podcasts and video, where word timestamps make editing and captioning easier
- Call-center analytics on recorded calls, with personal-information tagging as a first pass before analysis
- Research recordings, such as user interviews or field recordings in several languages
- Audio search and question answering, for example “what did the second speaker say about delivery dates?”, in English
If you want a finished meeting product rather than a model to build on, tools such as Granola, Otter.ai, and Fireflies.ai already handle capture, notes, and sharing.
Consent and Privacy
Self-hosting keeps audio on infrastructure you control, which helps with privacy rules and confidential material. It does not settle consent. Recording laws differ: some places require everyone in a conversation to agree, and workplace and call-center recordings often come with notice obligations. Tell people when they are being recorded and transcribed, store transcripts with the same care as the audio, and decide how long you keep both. The license itself puts responsibility for consents on you.
Two technical points also matter. Speaker labels separate voices within one recording (“Speaker 1,” “Speaker 2”); they do not identify real people, so matching labels to names is a separate step that deserves the same care as any identification. And the built-in personal-information tagging is a helper, not a guarantee: check what it misses before relying on it for redaction.
Limitations
- Recorded audio only. It is not a streaming model, so it does not suit live captions or voice agents.
- One hour per recording. NVIDIA does not guarantee results beyond that; split longer audio.
- English-only extras. Audio question answering and instruction following were evaluated only in English.
- Heavy hardware. Plan for a data-center GPU, a top-end workstation card, or a hosted provider.
- Uneven runtime support. Timestamps, speaker labels, and translation are not available in every runtime.
- Occasional repetition loops. NVIDIA notes rare generation loops and provides a post-processing guardrail.
- Benchmarks are mostly vendor-selected. Independent testing of the release is still early.
Alternatives and Related Reading
- Whisper (OpenAI): the long-standing open-source baseline, far smaller and easier to run, though without built-in speaker attribution
- Deepgram: API-first speech recognition with streaming and voice-agent features; see speech-to-text
- ElevenLabs: transcription alongside text-to-speech and voice agents on one platform
- Mistral’s Voxtral 3: batch and realtime transcription in public preview, closed-weight for now
- TwinMind vs Granola, for finished apps that capture meetings and conversations
Official sources: the model card on Hugging Face and NVIDIA’s launch post on Nemotron 3.5 Transcribe and Granary v2.
Continue learning
Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.
See how this tool fits into a workflow
Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.
Frequently Asked Questions
What is NVIDIA Nemotron 3.5 Transcribe?
It is an open-weight speech model from NVIDIA, released October 28, 2026. You give it a recording of up to about an hour, and it returns a transcript, optionally with word timestamps and speaker labels. The same model can also translate speech to or from English, identify the language, and answer questions about the audio. It works on recorded audio, not live streams.
Is Nemotron 3.5 Transcribe free to use commercially?
The weights are free to download, and NVIDIA's model card says the model is ready for commercial or non-commercial use under the OpenMDW-1.1 license. If you redistribute it, keep a copy of the license and its notices. Running it is not free: you pay for your own GPUs, or for a hosted provider such as Baseten, DeepInfra, or Simplismart.
Can I run Nemotron 3.5 Transcribe on my own computer?
Only on serious hardware. The model has about 34 billion parameters in total, and all of them have to sit in memory even though only about 4.3 billion are active for each token. NVIDIA recommends one H100 80GB GPU. GGML builds target an RTX 5090 or RTX PRO 6000, and compressed FP8 and NVFP4 versions reduce memory use, but a typical laptop will not run it.
Can it transcribe live meetings or phone calls in real time?
No. NVIDIA describes it as an offline model built for recorded audio, not live streaming. It suits processing a meeting or call after it ends. For live captions or voice agents, you need a streaming speech-to-text service.
What is Granary v2?
Granary v2 is NVIDIA's multilingual speech dataset and the curation pipeline that produced it: about 463,000 hours across 48 languages. It is one of the main sources of training data for Nemotron 3.5 Transcribe and can be used to fine-tune it, but it is a dataset, not a model or another name for the transcription model.
Is it the most accurate transcription model available?
NVIDIA reports that it ranks first among open-weight models on several leaderboards, including Hugging Face's English Open ASR Leaderboard. That is a strong result, but it is not the same as being the best model overall, and leaderboard audio is not your audio. Test it on a sample of your own recordings, with your accents, crosstalk, and vocabulary, before choosing it.
Last updated: