Nvidia Launches Nemotron 3.5 Transcribe AI Speech Model
Credit: Wikipedia
Nvidia has introduced Nemotron 3.5 Transcribe, a new open-weight AI model designed to convert speech into text, translate spoken languages and identify individual speakers in audio recordings. Released on October 28, the model aims to give developers more flexibility when building transcription tools, multilingual applications and AI-powered workplace solutions.
Multilingual Speech Recognition and Translation
Nemotron 3.5 Transcribe supports speech recognition in 40 languages and can translate 31 languages to or from English. According to Nvidia, it can process recordings of up to one hour and distinguish between multiple speakers, making it suitable for meeting transcription, interviews, podcasts and customer service recordings.
One of the model’s key features is its ability to adapt transcription style through natural-language instructions. Users can request verbatim transcripts, cleaner written text or specialized terminology for a particular industry without retraining the entire model.
Open Weights for Developers
Nvidia is releasing the model weights alongside tools for customization and fine-tuning through its NeMo Speech ecosystem. The model is available in several formats, including BF16, FP8 and NVFP4, giving developers options for different hardware configurations.
The underlying architecture combines a speech encoder with a mixture-of-experts language model containing approximately 34 billion parameters, of which around 4.3 billion are active during processing. Nvidia says this design helps balance transcription quality with computational efficiency.
The company also reports competitive results on public speech recognition benchmarks, although actual performance can vary depending on audio quality, language and recording conditions.
Potential Applications and Limitations
Nemotron 3.5 Transcribe could help developers build tools for meeting summaries, searchable audio archives, multilingual communication and specialized transcription workflows. Its speaker attribution and word-level timestamps may also be useful for editing recorded conversations and analyzing discussions.
However, the model is designed primarily for recorded audio rather than live, streaming speech. Organizations planning to deploy it will also need to consider hardware requirements, integration costs and the terms of its OpenMDW-1.1 license.
With this release, Nvidia is expanding its open-weight AI portfolio beyond text and reasoning models into speech processing. The new model offers developers another option for building customizable speech recognition systems without relying exclusively on proprietary transcription services.
