NVIDIA releases promptable Nemotron 3.5 Transcribe

NVIDIA's October 28 announcement presents a model for developers turning recordings into usable text. An instruction can request a transcript with speaker labels or ask what decisions a meeting reached. These are different outputs from the same model. It processes recorded audio; this is not live transcription. NVIDIA recommends checking names, speaker changes and answers against the recording.

The model documentation makes an important distinction for Hungarian users: transcription and translation between Hungarian and English are supported, but Hungarian multispeaker transcription is not specified. Audio understanding and reasoning were evaluated only in English and may be unreliable in other languages. Speaker tags distinguish voices within a recording; they do not identify real people. Plausible output can still misrepresent the audio, and supplied context can steer it toward incorrect terms.

An Agora classroom demo shows a concrete use. Mason Adams describes a live class carried by Agora and recorded in the browser. After the session ends, the demo sends recorded audio to an early-access version of the model, then displays its word-timed transcript beside the replay. Adams describes searching for a term and jumping to that moment as a potential learning aid. The demo has not been tested at the very large audience sizes discussed in his article.