NVIDIA releases Nemotron 3.5 Transcribe and Granary v2 for multilingual speech-to-text, speaker attribution, and translation
Open-weight model achieves top results on multiple speech benchmarks, with an accompanying open data pipeline and fine-tuning tools for domain adaptation.
1 source · cross-referenced
- Open-weight Nemotron 3.5 Transcribe supports transcription in 40 languages, translation between 31 of those and English, and speaker attribution in 13 languages.
- Ranks first among open-weight models on the Hugging Face Open ASR Leaderboard (English) with a 5.33% average word error rate.
- Granary v2 provides an open multilingual data pipeline and fine-tuning recipes to adapt the model to domain-specific vocabulary and accents.
- Quantized versions are planned for local deployment on NVIDIA RTX Spark, DGX Spark, and Jetson AGX Thor devices.
NVIDIA published Nemotron 3.5 Transcribe and Granary v2 on Hugging Face and ModelScope as open-weight resources for multilingual speech-to-text, speaker attribution, translation, and audio question answering. The model transcribes 40 languages, translates between 31 of those and English, and attributes speakers in 13 languages. It supports recordings up to one hour long and can label up to eight speakers.
Developers can guide output with natural language prompts and choose where to process audio. The model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English with a 5.33% average word error rate, first on the multilingual track with a 4.62% mean WER across seven languages, and first on the long-form track with a 10.98% mean WER. On Artificial Analysis’ speech-to-text leaderboard, it ties for first among open-weight models at 2.8% AA-WER with about 4 billion active parameters.
For speaker attribution, Nemotron 3.5 Transcribe posts the lowest diarization error rate on David AI’s 21-language DAI-ASR-I18N combined track and ranks first on VoiceArena’s English full-session benchmark with the lowest concatenated minimum-permutation WER among 14 systems.
The model uses a speech encoder paired with a hybrid Mamba–Transformer mixture-of-experts language model, activating roughly 4 billion of its 30 billion parameters per token. Quantized versions are planned to enable local deployment on NVIDIA RTX Spark, DGX Spark, or Jetson AGX Thor devices, and scaling on Blackwell data-center GPUs.
Granary v2 provides an open multilingual data pipeline and reusable curation tools on Hugging Face and ModelScope, alongside NeMo Speech fine-tuning tools. In a medical case study, fine-tuning with Granary v2 reduced word error rate on held-out speakers by 45.8% relative to the best prompt.
The model card includes presets for different tasks, a prompt guide, and a task-by-language matrix. Word-level timestamps are supported, and developers are advised to spot-check names, speaker changes, and answers against the recording.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79