Nvidia’s Nemotron 3.5 Transcribe targets recorded audio workflows
Illustration, not documentary evidence of the event.
Nvidia has released Nemotron 3.5 Transcribe, an open speech model that accepts text instructions for transcription, translation, speaker labels and questions about recorded audio, according to the-decoder.com, reporting Nvidia’s announcement. It handles recordings up to an hour long with up to eight speakers.
Nvidia says the model transcribes 40 languages, translates 31 to or from English, and identifies speakers in 13. Those different coverage levels matter for teams choosing a meeting or interview workflow: support for transcription in a language does not establish support for translation or speaker labels in that language.
The first deployment decision is whether the job can wait until recording ends. Transcribe supports recorded audio, not real-time processing. For archived conversations, combining transcription and speaker attribution could simplify a workflow that previously required separate models. For live conversations, Nvidia’s recently introduced Nemotron 3 Diarization supports speaker recognition but needs a separate transcription model. These releases therefore address different operating needs despite their overlapping speaker features.
A second decision is whether specialized vocabulary warrants fine-tuning. In Nvidia’s experiment with recordings of Indian doctors and medical students, a more precise prompt reduced word error rate from 30.8% to 23.8%; fine-tuning brought it to 11.8%. Calculated from those reported results, prompting removed 7 percentage points of error, while the subsequent tuning step removed another 12 points. Relative to the improved prompt, that second reduction was 12 ÷ 23.8 × 100, or about 50.4%. This is a comparison within Nvidia’s experiment, not an expected improvement for another team’s recordings.
That result supports a staged adaptation decision: establish a stronger prompt before deciding whether training is justified. Nvidia reports that tuning took about 14 GPU hours on a single GB300, but also slightly worsened general English accuracy, while a competing medical speech model remained more accurate on brand-name drugs. A team serving both specialist and general conversations should therefore compare the prompted base model and tuned version on both kinds of audio before replacing its default. Nvidia recommends reserving at least 100 real, verified recordings for custom-model testing; its own tradeoffs make vocabulary accuracy and general-speech regression concrete acceptance criteria.