NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
Source Entity
Hugging Face - Blog
Nvidia has launched Nemotron 3.5 Transcribe, an open-weight model capable of multilingual transcription, speaker diarization, and translation. Alongside the Granary v2 data pipeline, this release aims to enhance specialized applications in fields like medical and business documentation.
Nvidia Advances Speech AI with Nemotron 3.5 and Granary v2
Nvidia has officially entered the next phase of speech-to-text technology with the release of Nemotron 3.5 Transcribe, an open-weight model designed to tackle the complexities of multilingual audio processing. By integrating transcription, speaker attribution (diarization), translation, and audio question-answering into a single framework, Nvidia is addressing the fragmented nature of current voice-AI workflows. The model is now accessible via Hugging Face and ModelScope, signaling a push toward broader developer adoption.
Bridging the Gap in Professional Transcription
The primary challenge in modern transcription is moving beyond simple word-for-word capture. As noted in the release, basic transcripts often fail to identify speakers or handle domain-specific jargon—critical failures in sectors like healthcare or corporate law. Nemotron 3.5 seeks to solve this by providing a promptable interface that allows applications to adapt to specific contexts, such as distinguishing between speakers in a fast-paced meeting or correctly formatting medical terminology.
Technical Capabilities and Performance
Boasting support for up to 40 languages and the ability to distinguish up to eight distinct speakers, Nemotron 3.5 represents a significant benchmark achievement. According to Nvidia, the model has demonstrated top-tier performance on the Hugging Face Open ASR Leaderboard, outperforming existing competitors in accuracy and versatility. Furthermore, the model facilitates translation to and from English for 31 languages, handling audio recordings up to one hour in length, which makes it highly suitable for enterprise-grade batch processing.
The Role of Granary v2 in Model Customization
A pivotal component of this release is the Granary v2 data pipeline. By providing an open-source speech dataset alongside fine-tuning recipes, Nvidia empowers developers to tailor the model to niche requirements. This "open-weight" approach is essential for industries where proprietary jargon or regional dialects are common; developers can leverage the NeMo Speech model ecosystem to refine the AI’s performance for specific business or clinical environments.
Market Positioning and Future Implications
While Nemotron 3.5 offers robust features for recorded audio, it is important to note its current limitations. As an offline-focused tool, it does not support real-time, live-audio transcription, leaving that market segment to incumbents like ElevenLabs. However, by positioning itself as the premier solution for post-meeting analysis and medical documentation, Nvidia is carving out a defensive moat in the enterprise productivity space. As these models become more adaptable, we can expect a shift toward highly specialized, industry-specific AI agents that rely on these foundational transcription models to drive automation in the workplace.