NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026, giving media, education and customer-service teams an open transcription model for recordings in 40 languages.
It is a component for building recording-processing tools, not a finished meeting app. NVIDIA documents recordings up to one hour and up to eight speakers. It processes recorded audio rather than providing native live transcription.
A classroom example shows the distinction. Agora describes a demonstration that sends audio to the model after a class ends, then displays a transcript with word-level timing alongside the replay. That is a way to revisit a lesson, not live captioning.
Language coverage depends on the task
NVIDIA’s announcement and language-support matrix distinguish three capabilities:
Transcription: turning speech into text in 40 languages.
Translation: translating speech between English and 31 other languages, not every pairing among the 40.
Speaker-labelled transcription: separating different voices in 13 languages.
Hindi is listed for all three tasks. Bengali, Marathi, Tamil and Telugu are listed for transcription, but not translation or speaker-labelled output. The model card also says speaker labels distinguish voices within a recording; they do not establish anyone’s real-world identity.
Audio question answering has a narrower evidence base: NVIDIA evaluated audio understanding and reasoning only in English and warns that these functions may be unreliable in other languages. The 40-language headline is not a promise of multilingual meeting answers.
Prompts can request verbatim or cleaned-up text and supply relevant vocabulary. Removing fillers and repetitions changes the record, so teams should choose the output style deliberately. NVIDIA also describes Granary v2 data-curation resources and NeMo Speech training tools for further adaptation.
Running it yourself still needs substantial hardware
The main model files are publicly listed, and the OpenMDW-1.1 licence permits use without a licence fee, subject to its terms. Hosting and integration still cost money. NVIDIA’s documented reference setup uses one H100 GPU with 80 GB of memory—not an ordinary laptop. That is a supported example, not a universal hardware minimum.
Availability of the memory-saving FP8 and NVFP4 versions is unclear. The announcement calls them forthcoming, while the model card describes them as available. Public access to those variants remains unconfirmed.
Hosted access has its own feature limits
DeepInfra advertises a launch rate of US$0.002 per audio minute. For illustration, 100 hours of recordings would cost US$12 at that rate: 100 × 60 × US$0.002, before any other charges or application costs.
Its launch API offers transcription, timestamps, vocabulary prompts and translation into English, but speaker attribution is still forthcoming. The hosted interface therefore does not yet expose every capability of the underlying model.
The practical implication: a team needing speaker-separated Hindi meeting notes should not choose that service on the model’s language list alone. Hindi speaker labels are documented in the model, but not yet available through that API. Check the recording’s language, the required output and the host’s actual features together.
For a separate managed-service option, our coverage of Azure LLM Speech’s Indic-language expansion explains its transcription and translation boundaries.