NVIDIA's Nemotron 3.5 Audio Models: What Enterprise Teams Should Test

NVIDIA's Nemotron 3.5 Transcribe is an open-weight model that brings multilingual transcription, translation, speaker attribution, and audio question answering into one checkpoint. Granary v2, the accompanying dataset and curation pipeline, gives developers a path to adapt it. Both are worth evaluating for speech workflows. But benchmark results don't tell you how the model will handle your company's speakers, audio conditions, or terminology. ¹²

The question that matters isn't whether the capability list is long. It's whether the model improves quality and cost on a specific workflow enough to justify the testing and adaptation work needed to find out.

What NVIDIA released with Nemotron 3.5 Transcribe and Granary v2

Nemotron 3.5 Transcribe is an open-weight, promptable model that bundles several audio tasks into one checkpoint. NVIDIA reports transcription in 40 languages, translation in 31 languages to or from English, and speaker attribution in 13 languages. It also supports word-level timestamps, recordings up to an hour, and up to eight labeled speakers. ¹

Those numbers aren't interchangeable. NVIDIA says language coverage varies by task, and audio question answering was trained in English. If your use case needs transcription in one language and translation from another, verify each combination separately rather than treating the headline language count as blanket support. NVIDIA says a task-by-language matrix is available. ¹

Granary v2 is the data release behind the model. Its public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1. That describes the released dataset—not necessarily the model's full training mixture or how it was weighted during training. ²

What the WER benchmarks establish, and what they don't

NVIDIA's model card reports a 5.33% average word error rate (WER) on the English Hugging Face Open ASR Leaderboard, first among open-weight models in that comparison. It also reports 4.62% mean WER across seven languages and 10.98% on the long-form track. On Artificial Analysis's Speech-to-Text leaderboard, NVIDIA reports a tie for first among open-weight models at 2.8% AA-WER. On the AA-AgentTalk set, it reports 1.8% WER versus 2.3% for the next-best open-weight model. ¹

These numbers are useful for screening a shortlist, but they don't establish how the model will perform on your audio. The supplied material does not show results for a specific organization's vocabulary, accents, background noise, overlapping speakers, or recording setup. ¹² The same gap between headline claims and your own conditions shows up with other specialized releases; see LiquidAI's LFM3: Six Specialized Models, Benchmark Claims, and Evaluation Gaps for a parallel case.

There's also a discrepancy worth flagging on long audio. NVIDIA says the model supports recordings up to an hour; SiliconANGLE reports that it can process and answer questions about 100 hours of audio. Neither source explains whether that 100-hour figure means one continuous input, a batch job, or something else. Don't plan a long-form workflow around that number until it's clarified. ¹²

Analysis: Leaderboard scores are a reason to run a pilot, not a substitute for one. They can narrow the shortlist; representative internal audio is needed to assess whether the error rate is tolerable for the workflow you're building.

How to evaluate a multilingual speech transcription model

Build the test around what the business needs, not around the full feature list. If the workflow touches transcription, translation, speaker labels, timestamps, or audio Q&A, test each separately—a strong result on one task says nothing about the others. ¹

Use recordings that match the conditions your system will actually face, including domain terms, expected accents, background noise, overlapping speakers, and relevant recording setups. The published results don't establish performance under those specific conditions. ¹² Hold out a sample for final comparison so prompt or adaptation work isn't tuned against the same recordings used to judge it.

Define success per task. Track WER for transcription. For other tasks, set separate review criteria: do translations preserve meaning, do speaker labels land on the right voices, are timestamps usable, and are answers to audio questions correct? The sources describe capabilities and benchmark numbers, but don't provide a single shared metric across these tasks; you'll need to build your own rubric. ¹

Prompts deserve their own test track. NVIDIA says natural-language instructions can steer output and provide domain vocabulary, and that the same recording can be prompted toward different outputs. ¹ Compare a simple instruction with a structured prompt that specifies vocabulary, task, and output format. Keep the audio and review process identical between runs so any score change can be assessed consistently.

Test area	What to check
Language and task	Transcription, translation, speaker attribution, timestamps, and Q&A where relevant; verify each needed language-task pair. 1
Audio conditions	Domain vocabulary, accents, noise, overlapping speech, and the recording setups your workflow actually uses. 12
Output quality	WER for transcription, plus task-specific review for translation, speaker labels, timestamps, and answers. 1
Prompt behavior	Whether vocabulary and output instructions improve results on the same held-out audio. 1

When prompting is enough, and when AI transcription fine-tuning earns a test

Start with prompting as a low-cost baseline. Supply relevant vocabulary and output instructions, then compare results on the same held-out audio used for the unprompted run. NVIDIA describes prompting as a way to guide output and supply domain vocabulary. ¹

The published adaptation numbers show why fine-tuning may be worth testing, not what it will deliver for you. SiliconANGLE reports 30.8% WER out of the box and 20.8% with careful prompting. Separately, it reports that fine-tuning for specialized work on the Eka Care Indian medical speech benchmark reached 11.8% WER. The report doesn't clearly establish that the first two figures are from that same benchmark, so treat them as illustrative rather than a controlled comparison you can generalize from. ² Separately, an NVIDIA medical case study reports a 45.8% relative WER reduction on held-out speakers versus the best prompt, without stating the absolute WER. ¹

A practical sequence:

1. Baseline the default behavior.
2. Add domain vocabulary and specify task and output format.
3. If the prompted result still falls short, test fine-tuning.
4. Compare every stage on the same held-out speakers and conditions.

The sources don't quantify how much company-specific data, compute, or engineering time fine-tuning will take, or how reliably reported gains transfer beyond those evaluations. ¹² Weigh that cost against the accuracy gain: a quality improvement that requires substantial data preparation may not pay off for a low-volume workflow.

What to measure for latency and infrastructure

SiliconANGLE reports that Nemotron 3.5 Transcribe is a 30-billion-parameter mixture-of-experts model with 4 billion active parameters, and says it can run on a single NVIDIA GPU depending on compression level. ² That's useful context, but not a deployment profile you can borrow.

The concrete throughput figure is hardware-specific: SiliconANGLE says a Blackwell Server Edition GPU using NVFP4 precision and one-quarter model compression processed 100 hours of audio in under three minutes. ² Treat that as a result tied to that configuration, not a latency estimate for your deployment.

On your own hardware and precision setup, measure end-to-end latency, real-time factor, throughput, memory use, concurrency, and failure behavior. Log audio duration and output task for every run, and resolve what the 100-hour figure represents before designing around batch or long-form processing. ¹²

Analysis: Keep infrastructure validation separate from quality validation. A model can produce acceptable transcripts and still be the wrong fit if throughput or resource use doesn't match your load.

What Granary v2 adds to the adaptation decision

Granary v2 provides three transcript styles: one retains fillers, one uses a different number-formatting convention, and one removes fillers. That's a concrete choice worth testing against whatever downstream system consumes the transcript. ²

The release also includes a reusable curation pipeline, presented alongside NeMo Speech fine-tuning tools as an adaptation path. ¹² That's a starting point, not evidence that data preparation will be easy or that results will match NVIDIA's reported evaluations.

Before using the dataset or adapting the model, check licensing, privacy, retention, and deployment terms, and estimate how much company-specific data your workflow would need to collect. The supplied material does not specify those terms. ¹² If audio or transcripts will also pass through hosted models, an enterprise AI gateway that supports hard routing constraints on region, data policy, and allowed providers can help enforce those decisions in one place.

What a pilot needs to decide

Keep three questions separate:

• Capability: Does the model clear your quality bar for the specific languages and tasks in scope?
• Infrastructure: Does it meet latency, throughput, memory, and reliability requirements on your target hardware?
• Adaptation: Do prompts or fine-tuning improve results enough to justify the data and engineering cost?

The model is reported as available on Hugging Face and ModelScope, with a downloadable checkpoint noted by SiliconANGLE. ¹² Reproduce an evaluation on your own audio before making production design decisions. If you also want to compare the self-run checkpoint against hosted alternatives, the model catalog is one place to see which models are available to test.

Keep a log for every run: model version, prompt, language, audio condition, hardware, precision, output format, and whether the result came from prompting or fine-tuning. Set go/no-go thresholds against the actual workflow—acceptable transcription error, translation review burden, speaker-label accuracy, timestamp usefulness, latency, and adaptation effort—before you start.

Analysis: The right outcome might be a limited deployment, more adaptation work, or walking away. The published benchmarks earn the model a seat at the table; they don't answer the questions that decide the outcome: task-by-language coverage, noisy and overlapping speech, resource use on your hardware, transfer to new speakers, and the licensing terms attached to the model and dataset. ¹²

Sources

¹ “NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt,” Hugging Face, published 2026-10-28. https://huggingface.co/blog/nvidia/nemotron-transcribe-granary

² “Nvidia releases all-in-one adaptable Nemotron 3.5 audio transcription and translation model,” SiliconANGLE AI, published 2026-10-28. https://siliconangle.com/2026/10/28/nvidia-releases-all-in-one-adaptable-nemotron-3-5-audio-transcription-and-translation-model/