NVIDIA Unveils Nemotron 3.5 Transcribe and Granary v2 Open Multilingual Speech Recognition Stack

The Three-Question Framework

What Happened?

On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe alongside the Granary v2 open data curation pipeline. Nemotron 3.5 Transcribe is an open-weight, promptable multilingual speech-to-text model supporting transcription, translation, speaker attribution, and audio question answering across 40 languages. On the Hugging Face Open ASR Leaderboard, it achieved the #1 open-weight ranking with a 5.33% average English Word Error Rate (WER) and a 4.62% mean WER across seven evaluated global languages.

Why It Matters

Unlike static automated speech recognition systems that require distinct pipelines for transcription, speaker diarization, and domain post-processing, Transcribe responds to natural-language system instructions dynamically (such as filtering filler words, attributing speakers in overlapping dialogue, or adhering to medical terminology). Powered by vLLM inference and Sortformer acoustic speaker cues, the architecture transcribed 100 hours of continuous audio in under three minutes on a single NVIDIA GPU.

What Could Happen Next

NVIDIA has published the model checkpoint and interactive demonstrations on Hugging Face and ModelScope, alongside open data manifests for Granary v2. Inference integration is available immediately through vLLM, with quantized local deployment scripts for NeMo-Speech.cpp scheduled for subsequent release. Developers can run the modular four-stage Granary v2 pipeline using Ray and NVIDIA Dynamo backends to curate custom audio datasets.

Executive Key Takeaways

Nemotron 3.5 Transcribe takes #1 among open-weight models on Hugging Face Open ASR Leaderboard with 5.33% English WER and 4.62% multilingual mean WER.

Combines transcription, translation, speaker attribution, and audio Q&A across 40 languages in a single promptable checkpoint.

High throughput benchmark achieves 100 hours of audio transcription in under three minutes on a single NVIDIA GPU using vLLM.

Source: NVIDIA technical research blog published October 28, 2026, on Hugging Face, titled 'NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt'.

The architecture unifies acoustic representation learning with a conditioned language backbone by injecting Sortformer diarization cues directly alongside projected spectral embeddings. In addition to general speech recognition, the system captured top rankings on David AI's DAI-ASR-I18N multilingual diarization leaderboard and posted a record low 32.8% concatenated minimum permutation word error rate on VoiceArena's English conversational Diarization-Bench.

Granary v2, the data engine powering the model, introduces a four-stage pipeline encompassing Audio Preparation (metadata segmentation and acoustic tagging), Multi-Engine Transcription (routed primary and fallback ASR), Task Label Generation (character rate and hallucination filtering), and High Fidelity (HiFi) noise culling. In empirical testing on Eka Care's Indian doctor-patient dialogue benchmark, domain-specific prompt conditioning alone reduced medical terminology error rates from 30.8% to 23.4% without any parameter fine-tuning.

On general LibriSpeech test-clean benchmarks, model adaptation preserves baseline accuracy, recording a 1.69% WER. While deployment via vLLM supports high-concurrency cloud endpoints, memory footprints on consumer GPUs remain constrained pending the release of optimized NeMo-Speech.cpp 4-bit and 8-bit quantized checkpoints.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations