Nvidia has released Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription, translation and question answering that processes up to 100 hours of audio in a single pass. The 30-billion-parameter mixture-of-experts model uses 4 billion active parameters, supports 40 languages and labels up to eight speakers. For business users, the release matters because one adaptable checkpoint replaces a chain of speech detection, transcription and diarization tools.
Single-pass transcription in 40 languages
Traditional transcription systems detect speech, convert it to text, label speakers and then answer questions in separate stages. Transcribe combines those operations, accepting prompts such as "Transcribe the recording and label each speaker," "Translate this audio into English" or "What decisions were made in this meeting?" The same checkpoint produces different outputs depending on the prompt, including combined transcription and analysis tasks.
The model reports a 30.8% word error rate out of the box, falling to 20.8% with careful prompting, according to Nvidia. After domain fine-tuning on the Eka Care Indian medical speech benchmark, the error rate reached 11.8%. Users can also supply jargon or expected dialogue before transcription, which helps the model interpret specialized conversations more accurately.
Nvidia built the model with Granary v2, an open multilingual speech dataset created with NeMo Curator. Version 2 covers 843,579 hours across 48 languages and 19 source corpora, up from about 643,000 hours and 25 languages in the prior version. Each recording carries three transcript styles: full verbatim with fillers, normalized text with figures and symbols, and cleaned text without filler words.
What this means for corporate audio workflows
For companies working with meetings, calls and field recordings, a single prompt-controlled model simplifies deployment and maintenance. Small teams gain translation, speaker labeling and summaries without assembling several vendors, while large organizations can standardize one checkpoint across departments and adapt outputs by prompt. Availability as a Hugging Face demo with downloadable weights supports domain fine-tuning for legal, medical or customer-service vocabularies.
Hardware flexibility is a second practical effect, from Jetson AGX Thor and DGX Spark at the edge and desktop to Blackwell data-center GPUs. Nvidia said a Blackwell Server Edition GPU with NVFP4 precision and one-quarter compression processes 100 hours of audio in under three minutes. Buyers should still verify speed on their own hardware, language mix and audio quality, since compression and precision settings affect throughput.
The marker to watch is adoption of Granary v2 fine-tuning and published error rates on industry-specific datasets beyond the medical benchmark. If vendors and enterprise teams release adapted checkpoints with materially lower error rates for finance, support centers or multilingual meetings, the single-model approach will have proven its business value.
