NVIDIA releases Nemotron 3.5 Transcribe, an open speech model for 40 languages

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that handles transcription, speaker attribution, translation and audio question answering in one checkpoint [1]. It is available on Hugging Face and ModelScope, alongside version 2 of NVIDIA's open Granary speech dataset, the company said on 28 October [1][2].

What can Nemotron 3.5 Transcribe do?

NVIDIA's Nemotron 3.5 Transcribe transcribes 40 languages, translates 31 of them and attributes speakers in 13. NVIDIA says it handles recordings up to an hour long and labels up to eight speakers [1]. Developers can steer the output with natural-language prompts, for example asking it to label each speaker or translate a recording into English [1].

Language support differs by task, and audio question answering is trained in English only, according to the blog post [1]. TechCrunch notes that the model works on recorded audio, not livestreams [2].

How does it score on benchmarks?

NVIDIA says the model ranks first among open-weight models on the English Open ASR Leaderboard at 5.33% average WER. On the Artificial Analysis speech-to-text index, NVIDIA reports a 2.8% AA-WER, tying for first among open-weight models [1]. For speaker attribution, it claims the lowest diarization error rate of 14 systems on David AI's 21-language leaderboard [1].

Only about 4B of the model's 30B parameters are active per token, according to NVIDIA. The model pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, which NVIDIA says puts it on the accuracy-versus-active-parameters frontier [1].

Can developers adapt it?

NVIDIA released Granary v2 data, a curation pipeline and NeMo Speech fine-tuning recipes so developers can adapt the model to domain vocabulary and accents [1]. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt, NVIDIA says [1].

Quantized versions are coming soon to run locally on RTX Spark, DGX Spark or Jetson AGX Thor, according to NVIDIA [1].

Our read

Opinion: Folding transcription, speaker labels and translation into one open checkpoint is useful for anyone building meeting or dictation tools, and the fine-tuning recipe matters more than the leaderboard rank. The benchmark wins are NVIDIA's own reporting, so independent tests on messy real-world audio will decide how good it really is.

What we could not verify

• The benchmark results; they are reported by NVIDIA in its blog post, and we did not re-run them.
• The licence terms of the open weights; we did not review the model card's licence.
• When the quantized versions will ship; NVIDIA says only that they are coming soon.