01Key takeaway
Teams that want to self-host transcription for meetings or interviews now have another commercially usable option that can run on a single GPU, though it only handles prerecorded audio files.
02What happened
NVIDIA released the weights for its Nemotron 3.5 Transcribe speech model (@NVIDIAAI). According to the official blog post, it uses a mixture-of-experts architecture, with about 30B total parameters and about 4B active per run (the model page counts the audio encoder too and lists about 34B). It can process audio files up to one hour long. It transcribes 40 languages, translates between 31 of them and English, and can identify speakers in 13 languages, with up to eight speakers. Prompts can provide specialized terms or ask questions about the recording. The license is OpenMDW-1.1, and the model page says commercial use is allowed.
Evaluation group @ArtificialAnlys says it scored 2.8% on the group’s AA-WER word-error-rate metric, tying for first among open-weight speech-to-text models.
03What this means for you
The model page says this is an offline model, so it cannot provide real-time captions or power a voice customer service line. Its audio-understanding and speech-instruction-following capabilities have only been evaluated in English; the recommended deployment uses one H100 80GB. Chinese is on the language list, so test accuracy on your own recordings before using it for Chinese meeting notes. These are statements from NVIDIA and the evaluation group (🗣 Self-reported); we have not tested the model ourselves.
04Sources
- @NVIDIAAI: NVIDIA’s announcement of Nemotron 3.5 Transcribe
- NVIDIA official blog: parameter count, language coverage, translation and speaker labeling, evaluation result
- Model page: license, offline limitation, hardware recommendation, language list
- @ArtificialAnlys: third-party evaluation; tied for first among open-weight models on the AA-WER metric