Nvidia introduced Nemotron 3.5 Transcribe and Granary v2 models

Nvidia has released two specialized models: Nemotron 3.5 Transcribe for speech recognition and Granary v2 for data structuring. These new releases are aimed at creating high-precision transcription systems with the ability to adapt to specific domains and linguistic nuances.

The Nemotron 3.5 Transcribe model is designed for audio-to-text conversion with multilingual support. In addition to transcription, it supports translation and speaker recognition, accepting instructions in the form of text prompts. The model demonstrates leadership across a range of benchmarks while using only one-sixth of the active parameters of its closest open-source competitor. For training, Nvidia has also provided a dataset including over 800,000 hours of speech, along with tools for fine-tuning.

Meanwhile, Granary v2 acts as a post-processing tool: the model takes transcribed text as input and converts it into structured formats, such as JSON or SQL, using prompts to control the output schema.

Both models support integration into data processing pipelines, allowing for the automation of report generation, meeting minutes, and analytical summaries from audio streams. Developers can fine-tune the systems on their own datasets, which is critical for highly specialized industries such as medicine or law, where standard models often make errors with professional terminology.

For practitioners, this means the ability to deploy an "audio-to-structure" pipeline using Nvidia tools, which reduces the time spent on manual labeling and data post-processing. The models are available for review and testing via the Hugging Face platform, providing developers with flexible settings for adaptation to specific business tasks.