NVIDIA Nemotron 3.5 Transcribe and Granary v2: The Speech AI You Can Prompt, Adapt, and Actually Own

For the last few years, speech-to-text has mostly been a black box. You upload audio, you get words back, and you hope the model understood the accent, the jargon, and the noisy room. If it didn't, you were stuck. You could beg the vendor for a fix or build an expensive workaround around the edges.

NVIDIA's new release of Nemotron 3.5 Transcribe, shipped alongside Granary v2, points at a different future. The pitch is simple and, if it holds up, genuinely important: intelligent multilingual transcription you can prompt and adapt. Not just a model that listens. A model you can instruct.

That shift, from fixed pipeline to steerable system, is the real story here. It says a lot about where AI is heading next.

What NVIDIA Actually Shipped

The headline is a transcription model built for multilingual work. The word "multilingual" matters because most speech tools treat extra languages as an afterthought. You get strong English and then a long tail that quietly falls apart. NVIDIA is positioning Nemotron 3.5 Transcribe as multilingual by design rather than multilingual by checklist.

The second half of the headline is more interesting: you can prompt it. In practice, that means the model responds to instructions about the task at hand. You can tell it what kind of audio it is dealing with, what kind of output you want, and what context matters. That is a big departure from the classic "one model, one behavior" setup.

The third piece is adaptation. A model you can adapt is a model you can tune toward your own world, your vocabulary, your product names, your speakers, your workflows. Instead of bending your business around the model, you bend the model toward your business.

Granary v2 arrived in the same release wave, part of the same Nemotron family push. Taken together, the message is clear: NVIDIA is building a stack of customizable, promptable foundation models rather than a single monolithic product. Transcription is one room in a much bigger house.

Why "Promptable" Is a Bigger Deal Than It Sounds

Think about how transcription is used today. A call center records a customer conversation. A newsroom turns an interview into a draft. A hospital dictates clinical notes. A global team records a meeting with five accents and two languages in the same hour.

Each of those jobs has different rules. A legal transcript needs exact wording. A meeting summary needs clean, readable sentences. A medical note needs the right terminology and nothing invented. A media workflow needs timestamps and speaker labels.

Old-school transcription forces all of those jobs through the same pipe. Promptable transcription lets you describe the job in plain language and get behavior that fits. That is not a small convenience. It is the difference between a tool you tolerate and a tool you build on.

It also changes who can use the model. You no longer need a machine learning team to get custom behavior. A product manager who understands the workflow can write an instruction. That lowers the barrier from "we need six months and three engineers" to "we need an afternoon and a clear description of the task."

Multilingual Is Now Table Stakes, Not a Feature

Global business does not happen in one language. Support tickets arrive in Portuguese, sales calls happen in German, engineering standups drift between English and Hindi, and compliance recordings pile up in a dozen languages nobody on the team speaks fluently.

For years, companies solved this with a patchwork: one vendor for English, another for regional languages, a translation service on top, and a human reviewer to clean up the mess. Every layer added cost and lost meaning. Idioms died in translation. Names got mangled. Tone disappeared.

A multilingual model that can also be prompted and adapted starts to collapse that patchwork into one step. That is where the real savings live, not in the per-minute price of transcription, but in the number of handoffs you can delete.

Why adaptation beats brute force

There is a pattern in AI right now. The first wave tried to solve everything with a bigger general model. The second wave is about fitting the model to the specific job. Adaptation is the second wave.

It shows up everywhere. Teams fine-tune on their own data. They add retrieval so the model can see internal documents. They write instructions that encode company rules. Transcription is following the same path, and it is a healthy sign. It means the industry is moving past the demo phase and into the plumbing phase, where tools get reliable enough to trust with real work.

What This Means for Businesses

Here is where it gets practical. A promptable, adaptable transcription model changes the math on several common projects.

• Customer support: Record every call, transcribe in the customer's language, and tag issues automatically. Prompt the model to flag complaints, compliance risks, or cancellation signals in the same pass.
• Media and publishing: Turn raw interviews into clean, structured drafts. Prompt for house style, speaker labels, and quote-ready formatting instead of cleaning up machine output by hand.
• Healthcare and legal: Adapt the model to your own terminology so it stops guessing at drug names and case citations. Human review stays, but reviewers spend their time on judgment instead of corrections.
• Global meetings: One transcript, multiple languages, consistent speaker handling. Prompt the model to produce both a verbatim record and a short summary from the same audio.
• Accessibility: Live captions that understand local accents and technical vocabulary, which is where generic tools usually fail.

None of these are new ideas. What is new is that the customization step is no longer a research project. That is what turns a pilot into a rollout.

The Developer Angle: Ownership and Control

There is a quieter argument in this release, and it is about control. Models you can prompt and adapt tend to be models you can run closer to your own data. For teams handling sensitive audio, health records, legal proceedings, internal strategy calls, that matters more than any benchmark.

It also changes the build-versus-buy conversation. When a model can be tuned to your domain, the vendor lock-in risk drops. You are not just renting behavior. You are shaping it. Teams should treat that as a strategic asset and document their prompts and adaptation data the way they document code.

A short checklist before you adopt

• Test on your own audio, not demo audio. Real rooms are noisy, real speakers interrupt each other, and real vocabulary is weird.
• Measure what matters to you. Word error rate is fine, but track name accuracy, number accuracy, and how often a human has to fix something.
• Write prompts like specs. Be explicit about format, terminology, and what the model should do when audio is unclear.
• Keep a human in the loop for anything legal, medical, or financial until you have months of evidence.
• Watch for invented content. Any generative model can fill a gap with something that sounds right. Build checks for it.
• Plan for privacy. Decide where audio lives, how long it is kept, and who can access transcripts.

The Risks Nobody Should Skip

Promptable transcription is powerful, and powerful tools have sharp edges. The biggest is confident wrongness. A model asked to produce clean, readable output may smooth over a hesitation or a correction that mattered. In a customer dispute or a legal matter, that is not a formatting choice, it is a lost fact.

Accent and dialect bias is another. Multilingual models are only as good as their coverage, and coverage varies. Teams should test across the actual speakers they serve, not a general sample.

There is also the temptation to over-automate. A transcript that is 95% right still needs a human to catch the 5%, and the 5% is often exactly where the risk sits. The winning pattern is not "remove the human." It is "move the human to higher-value work."

What This Signals About the Future of AI

Zoom out and the pattern is bigger than speech. The AI market is splitting into two layers. One layer is the frontier, huge general models chasing the next capability jump. The other layer is the applied layer, focused, customizable models that do a specific job well and can be shaped to fit.

Nemotron 3.5 Transcribe and Granary v2 sit firmly in that second layer, and that is where most of the economic value will land over the next few years. Most companies do not need a model that can do everything. They need a model that can do their thing, reliably, in their languages, on their terms.

Expect three things to follow. First, prompting becomes a core business skill, not a developer trick. Second, adaptation data becomes a competitive moat, the team with the best-tuned model wins on quality, not on which vendor they picked. Third, speech becomes an interface, not just a record. Once transcription is accurate, cheap, and steerable, voice stops being something you archive and starts being something you query.

The organizations that move early will not just save money on transcripts. They will unlock a searchable, analyzable, multilingual record of everything their business says, and that is a very different kind of asset.

The Bottom Line

The shift from "transcription as a service" to "transcription you can prompt and adapt" is the kind of change that looks incremental in a press release and feels enormous in production. It moves control back to the people with the domain knowledge. It turns a fixed tool into a flexible one. And it makes multilingual work a default instead of a project.

The technology will not be perfect on day one, and teams should test hard before trusting it with anything that matters. But the direction is right. The future of AI is not one giant model that does everything for everyone. It is many adaptable models that each do one thing well, and let you tell them how.