NVIDIA Open-Sources 30B Speech Model Weights: 4.3B Active Parameters, 4.71% English WER
NVIDIA released Nemotron 3.5 Transcribe as an open-weight model. This 30B MoE speech LLM has only 4.3B active parameters, supports 40 languages, outputs speaker labels and word-level timestamps, and achieves an average 4.71% Word Error Rate (WER) on the English Open ASR Leaderboard.
On October 28, NVIDIA officially launched Nemotron 3.5 Transcribe. Per MarkTechPost's coverage, this is an open-weight large speech model built on a Mixture-of-Experts (MoE) architecture, with a total parameter count of 30B and 4.3B activated parameters per inference run. It supports transcription across 40 languages, outputs speaker diarization labels and word-level timestamps, and posts an average WER of 4.71% on the English Open ASR Leaderboard.
WER refers to Word Error Rate, a metric where lower values indicate better performance. The 4.71% result is a strong figure for English transcription, but note that this score only corresponds to the English benchmark. No itemized performance data for all 40 supported languages is available in public disclosures. When implementing multi-language production use cases, you still need to run validation tests with your own audio samples.
Speaker labels and word-level timestamps are far more practical than plain raw transcribed text. Meeting records require distinguishing between different speakers, subtitles need word-level alignment, and video editing depends on precise timestamps to make exact cuts. These features directly determine whether transcription outputs can be seamlessly integrated into formal workflows, or only serve as rough reference drafts.
The cost efficiency of the MoE architecture is also worth calculating. The 30B total parameter count does not mean all 30B parameters are called for every inference, and the 4.3B active parameter specification translates to a much narrower computational path. For scenarios such as long-form audio processing and bulk batch transcription, deployment costs are often far more important than minor decimal differences on leaderboards. NVIDIA's move to open the model weights has a clear strategic intent: to get more speech applications running on its own inference stack. As for whether it is the best model in its size class, independent third-party multi-language test results will provide far more reliable conclusions.
Source: MarkTechPost, by Asif Razzaq. Original article link: https://www.marktechpost.com/2026/10/28/nvidia-nemotron-3-5-transcribe-open-speech-llm/
发布时间: 2026-10-29 12:41