Nvidia releases open-weight speech model handling 40 languages, including Arabic

Summary

• Nvidia released Nemotron 3.5 Transcribe, an open-weight model transcribing 40 languages and translating 31.
• One checkpoint handles transcription, speaker labeling, translation and audio questions, with Arabic among training languages.
• Open weights let regional firms run Arabic transcription on their own servers without cloud subscriptions.

The latest

A single open-weight checkpoint can now transcribe 40 languages, label up to 8 speakers and translate across 31 language pairs, after Nvidia released Nemotron 3.5 Transcribe on October 28, 2026 alongside the second version of its Granary dataset. The company said the model posts the lowest average word error rate among open-weight systems on Hugging Face's English Open ASR leaderboard.

Details

• The architecture: Nvidia said the model runs roughly 30 billion total parameters — its Hugging Face page lists 34 billion — with about 4 billion active per token, pairing a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model. One checkpoint performs transcription, speaker diarization, translation and question-answering about a recording.
• Language coverage: According to the announcement, the model transcribes 40 languages, translates 31 to and from English, and separates speakers in 13 languages, handling recordings up to one hour. Audio question-answering was trained in English only, Nvidia said.
• Arabic data: The model's technical report, published October 28, 2026 on research.nvidia.com, lists Arabic in the training language table with roughly 16,400 unique hours, and sets the license as OpenMDW-1.1.
• Accuracy scores: Nvidia reported 5.33% average word error rate on the English Open ASR leaderboard, first among open-weight models; 4.62% on the seven-language multilingual track; and 10.98% on long-form audio. On Artificial Analysis it tied for first among open models at 2.8% AA-WER, with about 4 billion active parameters against 24 billion for the model it tied with.
• Speaker separation: On DAI-ASR-I18N covering 21 languages, Nvidia reported an 18.26% diarization error rate, the lowest among 14 systems tested. On the English Diarization-Bench it reported 32.4% cpWER, first of 14 systems, while dropping 8.32% of reference words against a median near 31%.
• Throughput claim: Nvidia said 100 hours of audio process in under 3 minutes on a single RTX PRO 6000 Blackwell GPU, using an NVFP4 quantized build on 30-second clips at high concurrency. That quantized version is described as forthcoming, with no release date given.
• The dataset: Granary v2 contains 463,062 hours across 48 languages drawn from 16 source collections, up from 25 languages in the first release. Each recording carries three text variants plus translations with quality scores; metadata is public on Hugging Face and ModelScope, while audio is downloaded from each source under its own license.
• Domain tuning: In an Eka Care test on recordings of doctors and medical students in India, Nvidia said error rate on 2,764 held-out speaker clips fell from 22.44% to 12.17%, a 45.8% relative drop, using about 14 GPU-hours on one GB300. Error on LibriSpeech test-clean rose slightly, from 1.54% to 1.69%.
• Access and limits: Checkpoints are on Hugging Face and ModelScope with day-one vLLM and SGLang-Omni support, plus hosted endpoints at Baseten, DeepInfra and Simplismart. Nvidia said the model is built for recordings rather than live streaming, that language support varies by task, and that names, speaker turns and answers should be checked against the audio.

Between the lines

The Arabic opportunity sits in the license and the hardware numbers rather than the leaderboard. With open weights under OpenMDW-1.1, roughly 16,400 hours of Arabic in training and 100 hours of audio claimed in under 3 minutes on one GPU, a regional institution can keep sensitive recordings on its own servers. The Eka Care result suggests dialect and vocabulary tuning is cheap at 14 GPU-hours.

What's next

Watch for the NVFP4 quantized build and local deployment via NeMo-Speech.cpp, both announced without dates, and for independent Arabic benchmarks testing dialect handling and speaker separation, which Nvidia supports in 13 languages.

Source: Nvidia (Hugging Face blog), research.nvidia.com technical report