Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, speaker attribution, translation and audio question answering in a single checkpoint. The model is available on Hugging Face and ModelScope. It transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13. It handles recordings up to an hour long and labels up to eight speakers. Developers steer its output with plain-language instructions, such as asking for speaker labels, an English translation, or answers to questions about a meeting. They can also supply domain terms and adapt recognition to specialised vocabulary and accents. Word-level timestamps are supported. Nvidia is also publishing Granary v2, a multilingual dataset with a reusable curation pipeline, alongside NeMo Speech fine-tuning tools for training the model on new speech.
On published benchmarks, the model leads among open-weight systems on several measures. It posts a 5.33% average word error rate on the English track of the Hugging Face Open ASR Leaderboard. On the Artificial Analysis speech-to-text index it ties for first among open-weight models at 2.8% AA-WER, using about 4 billion active parameters against 24 billion for the model it ties. It also reports the lowest error rate on David AI's 21-language diarization leaderboard and ranks first in VoiceArena's 14-system comparison. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt. The article is written by the vendor, and language support varies by task. Audio question answering is trained in English.
For Australian businesses and professionals, the near-term change is more about options than cost. Because the weights are open, the model can run on local hardware such as RTX Spark, DGX Spark or Jetson AGX Thor systems, which may suit organisations that want recordings kept onshore. Quantised versions are not yet released, though. Recording meetings, calls or patient consultations still engages the Privacy Act 1988, including the Australian Privacy Principles on collection, security and disclosure. In several states and territories, recording a private conversation without consent can breach listening device laws. Health information is sensitive information under the Act. Teams worth trying it should test it on their own local audio, including Australian accents, and check names, speaker changes and numbers before relying on the output.
Key Points
Nemotron 3.5 Transcribe is an open-weight model that handles transcription, speaker labelling, translation and audio Q&A in one checkpoint.
It covers 40 languages, translates 31 to or from English, attributes speakers in 13, and handles recordings up to an hour with up to eight speakers.
Vendor-reported results include 5.33% English word error rate on the Hugging Face Open ASR Leaderboard and a 2.8% AA-WER tie for first among open-weight models.
Granary v2 data and NeMo fine-tuning tools let teams adapt the model; one medical case cut word error rate by 45.8% versus the best prompt.
Quantised versions are still coming, so local deployment depends on hardware and timing, and recording consent and privacy obligations still apply in Australia.