NVIDIA has published Nemotron 3.5 Transcribe on Hugging Face, the newest version of its multilingual speech transcription model family. The release highlights one central feature: the ability to steer the model's behavior through text instructions rather than relying solely on raw, fixed transcription output. In practice, this would let users adapt the output to their context, for instance requesting specific punctuation styles, formatting conventions, or adjustments for domain-specific vocabulary.
Alongside the model, NVIDIA has also released Granary v2, an updated version of the speech dataset it uses to train and evaluate its automatic speech recognition systems. Datasets of this kind typically consist of hours of annotated audio recordings spanning multiple languages, a key ingredient for building transcription systems capable of generalizing beyond English. Releasing the model and the dataset together reflects an approach NVIDIA has increasingly favored, positioning its Nemotron lineup as an open alternative to proprietary transcription solutions.
Choosing Hugging Face as the distribution channel also signals an intent to reach developers and speech-processing researchers directly, allowing them to test, fine-tune, or integrate the model into their own pipelines. Promptable, multilingual transcription models fit into a broader industry trend in which audio architectures are gradually inheriting the control mechanisms originally developed for text-based large language models.
Without more detailed technical information on benchmarks or language coverage in this announcement, it remains difficult to assess exactly how Nemotron 3.5 Transcribe compares to competing solutions such as OpenAI's Whisper or Google's transcription models. The main significance of this release lies in NVIDIA's continued push to open up its speech technology stack, a segment where the company is working to strengthen its footprint beyond hardware.