NVIDIA has released Nemotron 3.5 Transcribe, an AI model designed to convert recorded speech into text while handling speaker identification, translation, contextual terminology and different transcription styles. Announced on October 28, 2026, the model targets developers building voice assistants, meeting transcription software, customer-support systems and multilingual speech applications.
The release stands out because NVIDIA combines several speech-processing capabilities in one model. Users can request verbatim transcripts, cleaned-up written text, speaker-attributed conversations or translations through natural-language instructions rather than training separate models for every output style. NVIDIA has also published downloadable model weights and a technical report, giving developers the option to run and adapt the model on supported infrastructure. NVIDIA Research
Key Takeaways
- New speech AI model: NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026.
- 40 languages: The model supports multilingual transcription, although individual features vary by language.
- Speaker identification: It can attribute speech to different speakers, including overlapping conversations, in supported settings.
- One-hour recordings: The model can process recordings up to one hour long.
- Promptable transcription: Users can specify verbatim, written, contextual, speaker-attributed or translated output.
- Open weights: NVIDIA provides model checkpoints under its OpenMDW-1.1 license.
- Reported benchmark results: NVIDIA reports a 4.71% mean word error rate on nine public English test sets and 4.62% on the multilingual Open ASR Leaderboard.
What Is NVIDIA Nemotron 3.5 Transcribe?
Nemotron 3.5 Transcribe is a speech-augmented language model, or SALM, developed to handle more than basic speech-to-text conversion.
Traditional automatic speech recognition systems primarily convert spoken words into written text. Additional processing may be necessary to identify speakers, add timestamps, translate the conversation or adapt terminology to a particular industry.
NVIDIA’s model combines these tasks in a single system. It can process recorded conversations and generate text according to the instructions provided by the user.
For example, a journalist could request a verbatim transcript that preserves the speaker’s original wording. A business could request a polished transcript with cleaner punctuation and written numbers. A researcher analyzing interviews could ask the system to identify individual speakers and attach timestamps to their words.
This flexibility is important because a transcript suitable for legal review may need to preserve pauses and exact wording, while a transcript intended for publication may benefit from cleaner formatting.
NVIDIA’s approach allows developers to choose the transcription style through instructions instead of maintaining a separate model for every format.
How the Model Works
Nemotron 3.5 Transcribe combines a Transformer-based speech encoder, a frozen Nemotron 3 diarization model and the Nemotron 3.5 Lightning 30B-A3B language model.
In simple terms, the speech encoder processes audio, the diarization component helps determine who is speaking, and the language model generates the requested text output.
The model uses a mixture-of-experts architecture. It contains approximately 34 billion parameters overall, with around 4.3 billion active for each token. This allows it to draw on a larger collection of learned parameters without activating all of them for every processing step.
NVIDIA also incorporates multi-token prediction techniques to accelerate inference. An optional DFlash2 draft model is available for speculative decoding, which can help speed up transcription when supported by the deployment setup.
The overall design aims to combine detailed speech understanding with efficient processing. Actual speed depends on the hardware, model checkpoint, audio characteristics and inference configuration.
Key Features of Nemotron 3.5 Transcribe
| Feature | What it does |
|---|---|
| Multilingual transcription | Converts speech into text across 40 languages |
| Speaker attribution | Identifies and labels speakers in supported languages and tasks |
| Word-level timestamps | Adds timing information to individual words when requested |
| Contextual transcription | Uses supplied terminology and expected names to improve recognition |
| Speech translation | Translates speech between English and supported languages |
| Verbatim transcription | Preserves spoken wording rather than automatically polishing it |
| Written-form transcription | Produces text formatted for readability |
| Long recordings | Supports recordings up to one hour in a single pass |
Not every feature is available in all 40 supported languages. NVIDIA reports speaker-attributed transcription in 13 languages and English-related translation in 31 languages.
The model can handle up to eight speakers in supported scenarios. This is useful for meetings, interviews and group discussions, where identifying the correct speaker is often as important as recognizing the words.
Why Speaker Identification Matters
A transcript can contain the correct words and still be difficult to use if it does not identify who said them.
In a business meeting, for example, an incorrect speaker label could make it appear that one employee approved a decision when another person actually made the statement. In a research interview, confusing speakers can undermine the value of the transcript.
Speaker diarization attempts to divide audio into segments associated with different speakers. Nemotron 3.5 Transcribe integrates speaker information into its speech-processing architecture to produce speaker-attributed text.
NVIDIA says the model achieves strong results on overlapping multi-speaker speech, including long meetings and calls. Its technical report describes performance that exceeds specialized multi-stage systems on certain speaker-attributed recognition tasks.
These results are relevant to noisy or fast-moving conversations, but accuracy can still vary when voices overlap heavily, recordings have poor audio quality or speakers sound similar.
NVIDIA’s Reported Accuracy Results
NVIDIA reports the following results in its technical report:
| Benchmark | Reported result | What it measures |
|---|---|---|
| English Open ASR Leaderboard | 4.71% mean word error rate | Average transcription errors across nine public English test sets |
| Multilingual Open ASR Leaderboard | 4.62% mean word error rate | Transcription performance across the leaderboard’s multilingual test sets |
| ContextASR-Bench | Sixfold reduction in named-entity errors | Effect of supplying contextual information about expected names and terms |
Word error rate measures the number of substitutions, deletions and insertions relative to a reference transcript. Lower values generally indicate better transcription accuracy.
These numbers are reported by NVIDIA and should be interpreted within their benchmark settings. Results on public datasets do not guarantee the same performance on regional accents, noisy telephone calls, technical discussions or every supported language. Developers should test the model on representative audio before using it in production.
What Developers Can Build With the Model
Nemotron 3.5 Transcribe could support several types of applications.
Meeting and interview transcription: Companies can generate transcripts with speaker labels and timestamps, making conversations easier to review and search.
Customer-service analytics: Contact centers can transcribe calls, identify speakers and adapt output to company-specific terminology. Additional systems would still be needed for tasks such as sentiment analysis or quality scoring.
Media production: Journalists, podcasters and video teams can create transcripts and translated text for recorded content.
Multilingual applications: Developers can build services that transcribe speech in supported languages and translate eligible recordings into English or from English into supported languages.
Specialized vocabulary: Organizations can supply expected names, product terms and technical expressions to help reduce recognition errors for domain-specific language.
The model can be integrated into larger pipelines that include storage, search, summarization and analytics. However, transcription accuracy and speaker attribution should be checked carefully before using outputs for legal, medical or other high-impact decisions.
Open Weights, Hardware and Licensing
NVIDIA has published the model on Hugging Face, along with different checkpoint variants.
The available releases include BF16, FP8 and NVFP4 versions, as well as GGUF files intended for NeMo-Speech.cpp. These options give developers choices based on memory capacity, hardware compatibility and inference requirements.
The full BF16 model is designed for substantial computing resources, with NVIDIA listing a single H100 80GB GPU as a reference configuration. Quantized checkpoints can reduce memory requirements, but supported hardware and performance vary by version.
The model is released under NVIDIA’s OpenMDW-1.1 license. Developers should read the license and model documentation to understand the applicable conditions before using the weights in commercial applications.
NVIDIA has also released the Granary v2 dataset associated with development of the model. Together, the model weights, technical report and dataset provide resources for developers and researchers interested in speech recognition and adaptation.
The Bigger Picture
Nemotron 3.5 Transcribe reflects a broader shift in speech AI, where developers increasingly want one model to handle transcription, speaker identification, translation and context-specific formatting. Combining these capabilities can simplify software architecture and reduce the need to connect several specialized systems.
The release also strengthens NVIDIA’s position beyond GPUs and AI infrastructure. By publishing open-weight models and supporting tools, the company is competing to become part of the software stack that developers use to build AI products.
The real test will be performance outside benchmark datasets. Accuracy across accents, languages, noisy recordings and overlapping speech will determine how useful the model becomes in practical applications. For developers seeking a downloadable speech model with flexible output formats, Nemotron 3.5 Transcribe is a significant new option to evaluate.
FAQs
1. What is NVIDIA Nemotron 3.5 Transcribe?
It is an open-weight AI model that converts audio into text and supports speaker attribution, contextual transcription, timestamps and translation in supported languages.
2. How many languages does Nemotron 3.5 Transcribe support?
It supports 40 languages for transcription. However, features such as speaker attribution and translation are available in fewer languages.
3. Can it identify different speakers in a recording?
Yes. The model supports speaker-attributed transcription and can handle up to eight speakers in supported scenarios. Results depend on audio quality and how much the speakers overlap.
4. Is Nemotron 3.5 Transcribe free to download?
NVIDIA publishes the model weights for download under the OpenMDW-1.1 license. Running the model may still incur hardware, cloud-computing or infrastructure costs.
5. How accurate is the model?
NVIDIA reports a 4.71% mean word error rate across nine public English test sets and 4.62% on the multilingual Open ASR Leaderboard. Actual accuracy will depend on the language, recording conditions and task.
Looking Ahead
Developers will likely evaluate Nemotron 3.5 Transcribe against existing speech-recognition services on their own recordings, especially for speaker attribution, specialized vocabulary and multilingual conversations. Its open-weight availability offers flexibility for teams that want more control over deployment, customization and data handling.
As speech models become more capable, the competition will increasingly focus on more than word recognition. Reliable speaker labels, contextual accuracy, translation quality, inference speed and deployment costs will shape which systems become practical choices for businesses. NVIDIA’s release brings these capabilities together in one model, giving developers a new foundation for building speech-powered applications.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
