NVIDIA releases Nemotron 3.5 Transcribe, an open multilingual speech model

Transcribe covers 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour long, labels up to eight speakers, and takes natural-language instructions that set the output, such as a translation or labeled speaker turns. It also supports word-level timestamps and runs on a single NVIDIA GPU, from DGX Spark to Blackwell data-center systems.

NVIDIA cites several benchmarks. On the Hugging Face Open ASR Leaderboard it ranks first among open-weight models for English at 5.33% average word error rate, and on Artificial Analysis it ties for first among open-weight models at 2.8% AA-WER with about 4B active parameters. On David AI's 21-language diarization leaderboard it posts the lowest diarization error rate of the 14 systems tested. The model is built for recorded audio rather than live streaming.

NVIDIA also released Granary v2, a multilingual speech dataset covering 463,062 hours across 48 languages, along with NeMo Speech fine-tuning tools. In a medical case study on Eka Care's benchmark, fine-tuning cut held-out word error rate from 22.44% to 12.17%, a 45.8% relative reduction, using about 14 GPU-hours on one NVIDIA GB300.

Developers building meeting, call or dictation pipelines can use one model for transcripts, speaker labels and translations instead of stitching together separate tools. Because the model targets recorded audio, it fits post-call and batch analysis rather than real-time voice agents. For domain jargon, the study suggests keeping at least 100 real, labeled clips to select checkpoints and checking the entity types your application depends on.