NVIDIA has put a new open-weight speech model on Hugging Face and ModelScope that tries to replace a stack of separate tools with one checkpoint. Nemotron 3.5 Transcribe, released on October 28, takes recorded audio and, depending on a plain-language instruction, returns a transcript, a translation, a transcript with speaker labels, or an answer to a question about the recording. Alongside it the company published Granary v2, the data-curation pipeline and public manifests behind part of the model's training material.
The model card lists transcription in 40 languages, translation to or from English for 31 of them and speaker attribution in 13. A single file can run to an hour, and the model separates up to eight voices. Under the hood a speech encoder feeds a hybrid Mamba and Transformer mixture-of-experts language model with roughly 30 billion parameters in total, of which about 4 billion are active for each token. NVIDIA ships BF16, FP8 and NVFP4 versions under the OpenMDW-1.1 license, which the card says permits commercial and non-commercial use.
Prompting is the central design idea. Users can ask for verbatim or cleaned-up text, request a translation, or supply a short list of expected terms, such as drug names in a clinical conversation, and tell the model to use them only when the audio supports them. Audio question answering was trained in English only, and NVIDIA's own guidance is to double-check proper names, speaker turns and any answers by listening back to the audio. The model also targets files rather than live streams, a limit TechCrunch singled out as the main drawback.
NVIDIA backs the release with a long list of leaderboard results, all reported by the company. It says the model is the top open-weight entry on the English track of the Hugging Face Open ASR Leaderboard at 5.33% average word error rate, ties for first among open models on Artificial Analysis's AA-WER index at 2.8%, and records the lowest diarization error rate, 18.26%, among 14 systems on David AI's 21-language test. The model card adds that some training data was licensed from David AI, while stating that the evaluation set itself was excluded.
Hardware claims follow the same pattern. NVIDIA says one GPU is enough, from the Jetson AGX Thor module and DGX Spark up to Blackwell data-center parts, and that an RTX PRO 6000 Blackwell Server Edition card using NVFP4 can work through 100 hours of 30-second clips in less than three minutes at high concurrency. Developers can serve the checkpoint with vLLM, run it locally through NeMo-Speech.cpp, or call hosted versions from Baseten, DeepInfra and Simplismart.
Granary v2 is the part aimed at teams that need to adapt the model. Built with NeMo Curator, the pipeline segments audio, routes each language to primary and fallback recognizers, generates several written forms of each transcript and filters out noisy or overlapping clips. Its public manifest describes 463,062 hours across 48 languages drawn from 16 corpora, against 25 languages in the first Granary. The source audio is not bundled: users must fetch it separately under each corpus's own license.
To show what adaptation buys, NVIDIA ran a medical case study on Eka Care's benchmark of recordings from Indian doctors and students. On speakers held out from tuning, fine-tuning lowered word error rate from 22.44% with the best prompt to 12.17%, a 45.8% relative cut, after about 14 GPU-hours on a single GB300. General English accuracy slipped slightly, with LibriSpeech test-clean error rising from 1.54% to 1.69%.
The launch lands in a crowded month for voice AI. TechCrunch noted that Mistral shipped Voxtral 3 earlier in October and that ElevenLabs and Salesforce also released voice models. Nemotron 3.5 Transcribe joins NVIDIA's smaller Parakeet models. Whether its leaderboard lead holds up will depend on outside testing, since every benchmark figure so far comes from NVIDIA's own materials.