NVIDIA has released Nemotron 3.5 Transcribe, an open-weight model that combines speech transcription, translation, speaker labeling and audio question-answering in one system. Developers can get it on Hugging Face and ModelScope, and use natural-language prompts to request different outputs from the same recording, such as a transcript with speaker labels or a translation into English.
The Hugging Face post says the model transcribes 40 languages, translates 31 to or from English, and labels speakers in 13 languages, with up to eight speakers identified. It also supports word-level timestamps. Audio question-answering is trained in English, and the post describes recordings of up to an hour; SiliconANGLE says it can digest and answer questions about 100 hours of audio. The two accounts do not clarify whether they mean different input limits or capabilities.
What do the benchmark results show?
Nemotron 3.5 is a 30-billion-parameter mixture-of-experts model, with about 4 billion parameters active at a time. NVIDIA’s Hugging Face post reports that it ranks first among open-weight models on the Hugging Face Open ASR Leaderboard’s English track, with 5.33% average word error rate (WER); lower WER means fewer transcription errors. The post also reports 4.62% mean WER across seven languages and 10.98% on the long-form track, the lowest results among open-weight models in its comparison.
The post reports results from other evaluations too: 2.8% AA-WER on Artificial Analysis’s speech-to-text leaderboard, tied for first among open-weight models, and the lowest diarization error rate on David AI’s 21-language benchmark. NVIDIA says fine-tuning for medical speech brought WER on the Eka Care Indian medical speech benchmark to 11.8%. These figures describe particular tests, not a guarantee of accuracy on every recording.
How can developers adapt it?
NVIDIA also released Granary v2, an open multilingual speech dataset and data-curation pipeline, alongside fine-tuning tools. It covers 843,579 hours across 48 languages and 19 source collections, compared with about 643,000 hours across 25 languages in the earlier version. Developers can use the data and tools to adapt the model to specialist vocabulary, accents or preferred transcript formats. In a medical case study, the Hugging Face post says fine-tuning lowered WER on speakers held out of training by 45.8% compared with the best prompt.
The model runs on a single NVIDIA GPU, from Jetson AGX Thor and desktop DGX Spark systems to data-center hardware. NVIDIA says a Blackwell Server Edition GPU processed 100 hours of audio in under three minutes using compressed model settings; the Hugging Face post specifies that this test used 30-second clips at high concurrency. The post says the model is designed for recorded audio, not live streaming, and recommends checking names, speaker changes and answers against the recording.
Why it matters
Developers building tools for meetings, classes or medical dictation can try one model for transcripts, speaker labels and translation, then fine-tune it for specialist terms. Its reported benchmark results do not establish how accurate it will be on a particular recording, so users still need to check important names and answers.
