NVIDIA Nemotron 3.5 Transcribe Launches with Prompt Control

NVIDIA has launched Nemotron 3.5 Transcribe and Granary v2, two speech-to-text models that accept natural language prompts to control transcription output. Officially released on October 28, 2026, these models extend the Nemotron 3.5 family beyond vision-language and reasoning tasks into adaptive audio processing, letting developers shape transcription behavior through conversational instructions rather than fixed configuration parameters.
Prompt-Driven Transcription Control
Nemotron 3.5 Transcribe interprets natural language prompts to adjust how it converts speech to text. Users can instruct the model to include or exclude punctuation, apply specific capitalization rules, expand abbreviations, or preserve technical terminology. This approach eliminates the need for separate fine-tuning pipelines when transcription requirements vary across contexts, making the model suitable for medical dictation, legal proceedings, customer service logs, and technical documentation.
Granary v2 serves as the updated audio encoder backbone, handling multilingual input across dozens of languages. The architecture supports streaming and batch processing modes, allowing real-time transcription for live conversations and high-throughput processing for archived audio libraries.
Adaptive Formatting and Domain Vocabulary
The models can adjust output formatting based on prompt instructions. A prompt such as "transcribe with speaker labels and timestamps" produces timestamped, multi-speaker output, while "transcribe as a single paragraph without punctuation" returns continuous text. Domain-specific vocabulary can be injected through prompts: "use medical terminology" or "preserve software code snippets as written" guide the model to recognize and retain specialized terms.
This flexibility addresses a common pain point in enterprise transcription workflows, where different departments require different formatting standards. Legal teams may need verbatim transcripts with filler words intact, while marketing teams prefer cleaned, summarized versions of the same audio.
Multilingual Support and Performance
Nemotron 3.5 Transcribe supports transcription across more than 50 languages, with Granary v2 encoding audio features that preserve linguistic nuances such as tone, accent, and code-switching. The models handle mixed-language conversations, maintaining context when speakers alternate between languages mid-sentence.
NVIDIA has made the models available on Hugging Face, where developers can test prompts and evaluate performance against existing speech-to-text solutions. Early benchmarks show competitive word error rates on standard datasets, with the added benefit of prompt-driven customization that traditional models lack.
Integration with Nemotron Ecosystem
The transcription models join other NVIDIA Nemotron 3.5 releases, including the Bolt 74B reasoning model and Super VL 120B vision-language system. Developers building multimodal applications can now route audio through Transcribe, images through Super VL, and reasoning tasks through Bolt within a unified architecture.
NVIDIA positions these models as part of its broader strategy to provide enterprise-grade AI building blocks. The company has emphasized open access through Hugging Face while offering optimized inference options through NVIDIA AI Enterprise for production deployments.
What This Means
Nemotron 3.5 Transcribe and Granary v2 represent a shift from fixed-function transcription to instruction-driven audio processing. By accepting natural language prompts, these models reduce the engineering overhead required to adapt speech-to-text systems for specialized use cases. Enterprises can deploy a single model across multiple departments, using prompts to meet varying transcription standards without maintaining separate model versions. As prompt-controlled AI becomes standard across modalities, NVIDIA's approach suggests a future where customization happens at inference time rather than through pre-deployment fine-tuning.
on Emergent today





