NVIDIA Nemotron 3.5 Transcribe Brings Local Speech-to-Text

NVIDIA has launched a new model that turns audio recordings into text, identifies who is speaking and translates speech directly on the user's own machine. Called Nemotron 3.5 Transcribe, it lets developers build apps for meetings, customer service and voice work without always sending recordings to cloud servers.

According to NVIDIA's blog, the model supports speech transcription in 40 languages. Users can give it instructions in plain language: transcribe a recording and label each speaker, translate it, or pull the decisions out of a meeting. Developers can also add specialist terms to improve handling of vocabulary in medicine and healthcare, financial services and customer support.

Why local processing matters

Running the model on the device gives developers more control over where data lives. That is important when a conversation contains business information or sensitive details, and it reduces the need to move files to the cloud. Developers can run the model on supported hardware and later scale out to data centers when they need more capacity.

The release comes as NVIDIA expands the capabilities of its RTX Spark computers, which are designed to run AI models locally.

Sharing GPU power with background agents

NVIDIA also added tools that let users decide how GPU power is split between the application open in front of them and an AI agent working in the background. The main program can be given higher priority, or the agent can receive more resources to finish its task faster.

The change reflects the company's expectation that AI applications and agents will run continuously alongside traditional software on the same computer.