Product Launches
NVIDIA Releases Free Nemotron 3.5 Transcribe Model Covering 40 Languages and Eight Speakers
By News Desk ·
Key takeaways
- NVIDIA released Nemotron 3.5 Transcribe, a free open-weight model that transcribes audio in up to 40 languages and translates 31 languages to and from English.
- The model handles recordings up to an hour long, separates up to eight speakers and streams with latency as low as 0.07 seconds.
- NVIDIA paired the release with version 2 of Granary, its open-source speech dataset, on Hugging Face under the OpenMDW-1.1 license.
NVIDIA released Nemotron 3.5 Transcribe, a free open-weight speech model that transcribes audio in up to 40 languages and translates 31 languages to and from English, on October 28, 2026. Crypto Briefing reported that the model is available on Hugging Face under the OpenMDW-1.1 license, alongside version 2 of Granary, the company's open-source speech dataset.
Inside the Nemotron 3.5 Transcribe Release
The model transcribes speech across 40 language locales and handles recordings up to an hour long, according to research findings cited in the release. It separates up to eight speakers and supports real-time streaming built on a cache-aware FastConformer-RNNT architecture, reaching latency as low as 0.07 seconds to the final transcript. Developers can configure chunk sizes, trading speed for accuracy or the reverse, and punctuation and capitalization come built in.
Whalesbook described the model as a 30-billion parameter open-weight tool aimed at enterprise use in clinical and corporate settings, limited to pre-recorded audio rather than real-time transcription. The research findings cited by Crypto Briefing describe a model with 600 million parameters.
Granary first launched in August 2025 with roughly one million hours of audio data covering 25 European languages, according to the research findings. Version 2 supports the expanded multilingual model. The previous Nemotron 3 model was English-only, per the same findings.
NVIDIA's Speech Push in Company Records
NVIDIA's speech work sits alongside a broader AI platform build-out, according to akta.pro company records. The company reported first-quarter fiscal 2027 revenue of $81.6 billion, up 85% year-over-year, including $75.2 billion in data center revenue, up 92%. In June 2026 it issued $25 billion in high-grade bonds to expand liquidity and fund a 2,400% dividend increase alongside a commitment to return more than 50% of free cash flow to shareholders.
Earlier Steps Toward Open Speech Models
NVIDIA's Nemotron 3.5 Bolt E4B ran on an 8-vCPU Xeon VM with 16GB RAM and no GPU, correctly picking tools 60 out of 60 times and filling arguments 95% of the time, completing a full voice turn in about 4.5 seconds, HackerNoon reported in October 2026. Singtel's RE:AI sovereign AI cloud was preparing to deploy NVIDIA DGX GB300 systems in Asia-Pacific, which it described as the first deployment of the platform in the region, Telecompaper reported in October 2026. HIVE Digital Technologies secured approximately $129 million in three-year, 9% fixed-rate financing to purchase NVIDIA GB200 NVL72 systems for a three-year Cohere contract, TradingView reported in October 2026.