Nvidia Unveils Open-Source Nemotron 3.5 Transcribe and Granary v2 Speech Dataset Amid Booming AI Voice Tool Launches

Official Wednesday Announcement

Nvidia rolled out two new speech-focused AI assets this week: Nemotron 3.5 Transcribe, a fully free open-weight automatic speech recognition (ASR) model, and Version 2 of Granary, the company’s curated open-source speech dataset.

Core Capability Highlights

Nvidia’s official product specs position Nemotron 3.5 Transcribe as a highly capable ASR solution that supports direct audio transcription for up to 40 distinct languages, plus seamless bidirectional translation between English and 31 additional languages. The model is optimized to process full hour-long pre-recorded audio clips, and comes with built-in speaker diarization that can reliably distinguish up to 8 separate speakers in a single recording.

The firm claims the new model outperforms competing ASR offerings across all major industry benchmarks, and has secured a leading ranking on Hugging Face’s public Open ASR Leaderboard.

Target Use Cases and Ecosystem Support

Both new releases are purpose-built for developers building production AI tools for two of the most mature, widely adopted workplace AI use cases: corporate meeting transcription and medical clinical note documentation.

Developers can combine the 30-billion-parameter Nemotron 3.5 Transcribe model, the expanded v2 Granary speech dataset, and Nvidia’s existing NeMo Speech development toolkit to fine-tune custom ASR workflows that accurately recognize niche, industry-specific jargon across every supported language.

The only notable current limitation of the new Transcribe model is that it does not support real-time livestream audio processing, and works exclusively with pre-recorded audio files.

Cross-Hardware Deployment Flexibility

Despite its 30-billion parameter count, Nvidia confirms the model is optimized to run natively on a single Nvidia chip across the company’s full hardware stack: support spans edge devices including the Jetson AGX Thor robotics controller, DGX Spark chips powering next-gen AI laptops, all the way up to top-tier Blackwell architecture datacenter GPUs.

Nemotron 3.5 Transcribe is the latest addition to Nvidia’s growing family of open-weight voice AI models, which already includes the ultra-compact 1.1-billion-parameter Parakeet, a tiny but high-performance ASR model optimized for low-resource edge deployments.

Context of the Broader Industry Release Wave

Nvidia’s new speech tool launch lands in an unusually busy stretch of major AI voice product rollouts across the sector: earlier this month, Mistral AI released its Voxtral 3 speech model, ElevenLabs launched its low-latency Realtime speech engine purpose-built for instant, natural speech generation in conversational voice agents, and Salesforce debuted its own full-featured voice model, Launa-Omni.