Nvidia Releases Nemotron 3.5 Transcribe
Nvidia releases an open-weight transcription model supporting multilingual speech, speaker separation and related developer workloads.
Voice is becoming an interface for software agents, meetings and professional documentation. Nvidia is adding another open-weight option for developers building those systems.
What happened
Nvidia released Nemotron 3.5 Transcribe, a speech model that supports transcription across up to 40 languages. The company says it can translate among supported languages, identify multiple speakers, add timestamps and handle recordings lasting up to an hour.
Alongside the model, Nvidia updated its multilingual Granary speech dataset and described tools for customising transcription to specialist terminology. The model processes recorded audio rather than providing true live-streaming transcription at launch.
Nvidia reports strong benchmark results, but those scores do not guarantee comparable performance on noisy, accented or specialist recordings. Developers must also account for compute and memory requirements when selecting deployment hardware.
Why it matters
Speech-to-text software underpins medical notes, meeting records, customer support and voice interfaces. An open-weight model can give developers more control over deployment, customisation and handling of sensitive recordings compared with a strictly hosted API.
However, speech systems can still mishear names, technical terms or important details. In high-stakes settings, transcription accuracy and appropriate review remain essential.
The bigger picture
AI competition increasingly extends to specialist models and datasets that support applications built above them. Nvidia's release may strengthen its position in the development ecosystem while giving builders another way to integrate voice into products.