Nvidia releases open speech model for recorded audio

Nvidia released Nemotron 3.5 Transcribe on October 28, presenting it as an open-weight model for transcribing and analyzing recorded speech. The model can also translate audio, label speakers and answer questions about a recording, but it is designed for finished recordings rather than live transcription. Nvidia says the model handles 40 languages, translates 31 to or from English and can distinguish up to eight speakers in recordings of up to an hour. Users can guide its output with plain-language prompts, requesting, for example, a verbatim transcript, speaker labels, a translation or attention to specific terms. The same model can respond to questions about the recording’s content. The system has about 30 billion parameters, with roughly 4 billion active during each processing step, according to Nvidia. The company says it can run on a single Nvidia GPU, with supported hardware ranging from edge devices to data-center systems. The release also includes tools and recipes for adapting the model to specialized uses. Nvidia says its public Granary v2 dataset contains 463,062 hours of speech across 48 languages. The dataset and its accompanying data-building tools are intended...