Nvidia Unveils Open Multilingual Speech Model and Dataset

WHO:

SiliconANGLE - Nvidia Corp. released Nemotron 3.5 Transcribe and Granary v2.

TechCrunch - Nvidia released a free speech-to-text model; the article identifies ElevenLabs as a rival in real-time audio use.

MarkTechPost - NVIDIA released Nemotron 3.5 Transcribe, described as an open-weights speech large language model.

WHAT:

SiliconANGLE - Nvidia debuted Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription and translation, and Granary v2, an open-source speech dataset.

SiliconANGLE - The article describes Nemotron 3.5 Transcribe as a leaderboard-winning open artificial intelligence model.

TechCrunch - The model can transcribe up to 40 languages.

TechCrunch - The model works on recordings, not live audio; the article says it could support meeting and medical transcription tools, while real-time use remains an area for rivals such as ElevenLabs.

MarkTechPost - Nemotron 3.5 Transcribe has 4.3 billion active parameters and transcribes 40 languages.

MarkTechPost - The model provides speaker labels and word timestamps.

MarkTechPost - It posted a 4.71% mean word error rate (WER) on the English Open ASR Leaderboard.

WHERE:

MarkTechPost - The reported leaderboard result was on the English Open ASR Leaderboard.

HOW:

MarkTechPost - Nemotron 3.5 Transcribe is a 30-billion-parameter mixture-of-experts (MoE) model with approximately 4 billion active parameters; the article specifies 4.3 billion active parameters.

TechCrunch - The model processes recorded audio rather than live audio.

(UPDATED October 29, 2026 05:59 UTC)

Nvidia has released Nemotron 3.5 Transcribe, an open-weights model that can transcribe and translate recorded audio in 40 languages. It includes speaker labels and word-level timestamps, and Nvidia says it achieved a 4.71% mean word error rate on the English Open ASR Leaderboard. Nvidia also released Granary v2, an open-source speech dataset; the model handles recordings rather than live audio.