NVIDIA has released Nemotron 3.5 Transcribe, an open-weight speech model that turns recorded audio into a transcript, speaker labels, a translation or an answer to a question about the recording, all from one checkpoint. In its Hugging Face post, NVIDIA says the model ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English speech recognition, at a 5.33% average word error rate.
One checkpoint transcribes 40 languages, translates 31 of them and labels speakers in 13
NVIDIA says the model transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13. It handles recordings up to an hour long with up to eight speakers, and the output is steered with plain-language prompts such as “Transcribe this recording and label each speaker.” Audio question answering is trained in English only, so the language count depends on the task.
About 4B of its 30B parameters are active per token, and it ties a model with about 24B for first on a second board
The model pairs a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model, so about 4B of its 30B parameters are active per token. On the Artificial Analysis speech-to-text board it ties for first among open-weight models at 2.8% word error, tied with a model that has about 24B active parameters. NVIDIA’s blog post for RTX Spark laptops, published the same day, puts it second among open-weight models on that board, so the two NVIDIA documents describe the same result differently.
Running it locally on RTX Spark waits for quantized builds and NeMo-Speech.cpp
The RTX Spark post lists the model as available on RTX Spark. The Hugging Face post is more specific: the checkpoint available today is served with vLLM or SGLang-Omni, while local deployment with NeMo-Speech.cpp is coming soon alongside the quantized versions. The model is built for recorded audio, not live streaming.
The one throughput figure comes from a build that is not released yet. NVIDIA reports that the forthcoming NVFP4 quantized version processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, measured on 30-second clips at high concurrency. NVIDIA gives no speed figure for the checkpoint people can download today, and no figure for an RTX Spark laptop.