Nvidia Nemotron 3.5 Transcribe: open speech AI, 40 languages

Transcribe is promptable: the same recording can produce speaker-labeled transcripts, an English translation, or a summary of decisions made in a meeting, depending on the instruction. Developers can also feed in domain vocabulary, such as drug names, as hints, and Nvidia says the model should only use them when the audio supports it.

Nvidia tested targeted fine-tuning: on the Eka Care Indian medical speech benchmark, the word error rate dropped from 30.8% out of the box to 11.8% after adaptation, and 45.8% relative improvement on held-out speakers. Nvidia says the full fine-tuning workflow took about 14 GPU hours on a single GB300.

• About 844,000 hours of speech across 48 languages and 16 source corpora, up from roughly 643,000 hours in 25 languages in Granary v1
• Multiple transcript styles per recording: verbatim with filler words and spelled-out numbers, cleaned with digits, or filler-free
• Built with Nvidia NeMo Curator and shared so developers can run the pipeline on their own recordings
• Nvidia notes the public manifest does not fully match Transcribe's training mixture or sampling weights

For Tunisian developers and AI builders, the practical draw is access: open weights and an open dataset mean the model can be run locally or fine-tuned on Arabic and French audio without paying per-minute API fees, as long as you have one compatible Nvidia GPU. Nvidia has published fine-tuning guides and an extension that lets an AI coding assistant run the training workflow.

Two caveats matter. Transcribe handles recorded audio only, so live transcription still needs another tool, and Nvidia says audio question answering is trained in English only, with language support varying by task. The company also advises spot-checking names, speaker changes and answers against the source audio.

Nvidia has not announced pricing for any hosted service or confirmed availability in Tunisia; the model and dataset are free to download on Hugging Face. Anyone building multilingual or medical transcription products should look at whether the fine-tuning gains justify the extra engineering work.