Catch up on the essentials
- NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model designed for multilingual transcription, speech translation, speaker attribution and audio question answering.
- Natural-language prompts can request verbatim or cleaned-up transcripts, supply context such as product or medical vocabulary, request speaker tags and word-level timestamps, and direct speech translation.
- On ContextASR-Bench, NVIDIA reported a six-fold reduction in named-entity errors when the model received a list of expected entities, though that condition is not established as typical customer usage.
Selected from this article · 2026-10-29
Read on for the full pictureNVIDIA releases open-weight transcription model
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model designed for multilingual transcription, speech translation, speaker attribution and audio question answering. The model card lists its release across Hugging Face, GitHub through NVIDIA NeMo Speech, NGC and Build.Nvidia.com, with weights available under the OpenMDW-1.1 license.
The model supports 40 languages, according to NVIDIA. It can translate speech to or from English in 31 languages and attribute speakers in 13, with capacity for recordings up to one hour and labeling of up to eight speakers. Individual capabilities vary by language and task, and audio understanding and spoken-instruction following were evaluated only in English.
Prompts control transcript output
Natural-language prompts can request verbatim or cleaned-up transcripts, supply context such as product or medical vocabulary, request speaker tags and word-level timestamps, and direct speech translation. The model can also answer questions about audio, although that capability is trained in English.
Nemotron 3.5 Transcribe combines a Transformer speech encoder, a speaker-diarization model and NVIDIA’s Nemotron 3.5 Lightning language model. Its released checkpoint is designated 30B-A4B, with approximately 4.3 billion active parameters. NVIDIA released BF16, FP8 and NVFP4 checkpoints, along with GGUF files for NeMo-Speech.cpp and a fine-tuning toolkit through NeMo Speech.
NVIDIA reports benchmark results
NVIDIA reported a 4.71% mean word error rate across nine public English test sets on the Open ASR Leaderboard. It also reported a 4.62% mean word error rate across seven languages on the multilingual leaderboard and a 10.98% long-form result.
The company said the model ranked first among open-weight models in its leaderboard comparisons. Those rankings were based on the model card’s captured October 12, 2026 results and excluded API models. On ContextASR-Bench, NVIDIA reported a six-fold reduction in named-entity errors when the model received a list of expected entities, though that condition is not established as typical customer usage.
Optimized inference outpaces the released setup
NVIDIA reported throughput of up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell GPU using an NVFP4 checkpoint, DFlash2 speculative decoding and Nemotron 3.5 Transcribe inference software. The company said the release-day vLLM implementation was expected to run at roughly half the speed of that optimized configuration.
Actual performance depends on hardware, configuration and workload. NVIDIA also cautioned that Nemotron 3.5 Transcribe is an offline model: its internal audio chunking does not make it a real-time streaming service.
A related dataset accompanies the model
NVIDIA also released Granary v2, an open speech dataset and curation pipeline whose public manifest covers 463,062 hours across 48 languages and 16 source corpora. NVIDIA said those figures describe the dataset release, not the model’s complete training mixture. Built with NVIDIA NeMo Curator, the pipeline produces multiple transcript forms along with captions, speech-grounded question-answer pairs, contextual terms and translations with quality scores.
Share this article
