NVIDIA released Nemotron 3.5 Transcribe on Wednesday, October 28, 2026, an open-weight speech recognition model with a 30-billion-parameter mixture-of-experts (MoE) architecture — only 4.3 billion of which are active per inference step. The model combines verbatim transcription, translation to and from 31 languages, speaker attribution (up to eight distinct voices), and word-level timestamps, all within a single checkpoint that can fit on a consumer-grade GPU like an RTX 5090.
What distinguishes Nemotron 3.5 Transcribe from conventional automatic speech recognition systems is its ability to receive natural language instructions before the audio even begins playing. You can ask the model to transcribe a meeting recording literally, to summarize the key points, or to format the output in a specific way — the model adapts its output to the prompt without requiring fine-tuning. This approach, which NVIDIA calls "intelligent transcription," transforms a static speech recognition problem into a flexible interface where the same checkpoint produces different outputs depending on context.
On public benchmarks, the model leads the Hugging Face Open ASR Leaderboard for open-weight models with an average word error rate (WER) of 4.71% across nine English test sets, and 4.62% on the multilingual leaderboard. On the independent Artificial Analysis platform, which measures AA-WER (Adjusted Word Error Rate), the model scores 2.81% — second among all open-weight models, just 0.04 points behind leader Voxtral Small by Mistral, which uses roughly six times as many active parameters (24 billion versus 4.3 billion).
Efficiency is the most impressive feature. To achieve performance comparable to much larger models, NVIDIA employed a combination of Mamba-Transformer architecture with granular mixture-of-experts routing, where only a fraction of the parameters are activated for each token. The model was trained on Granary-v2, a multilingual speech dataset of 844,000 hours of speech across 48 languages — nearly a million hours — with varied transcriptions covering different accents, registers, and domains. Most of the training data is synthetic, which raises an important question about the model's generalization to real-world situations.
The launch comes amid an intense race in the voice AI sector. Mistral released the Voxtral family in early October, ElevenLabs launched its Realtime for instant voice agents, and Salesforce released the Launa-Omni. NVIDIA positions Transcribe as a lighter, more accessible alternative, especially for developers who want to run high-quality transcription locally — whether on a laptop with a Jetson AGX Thor, a DGX Spark, or a datacenter server with Blackwell GPUs.
But the model is not perfect. Transcribe works only with recorded audio, not live streaming — a significant limitation for real-time voice agent applications. Additionally, domain-specific benchmarks show that WER drops to 30.8% on domain-specific content before careful prompting, falling to 20.8% after. That means for medical, legal, or technical use, the model requires additional adaptation and human verification. NVIDIA itself recommends that users spot-check names, speaker changes, and model answers.
The weights are available under the OpenMDW-1.1 license, which permits commercial use, modification, and redistribution — though the Open Source Initiative is still evaluating whether the license formally meets the open source definition. Model weights are available on Hugging Face and ModelScope, with deployment support via vLLM and NVIDIA NeMo Speech. An interactive demo is available on Hugging Face Spaces to try the model before deploying to production.
The question that remains is whether Transcribe represents the future of local transcription — a model capable of running on any person's laptop, in dozens of languages, without depending on cloud services from companies like Google, Apple, or Microsoft — or whether domain-specific limitations and the absence of real-time streaming will be barriers sufficient for it to remain a niche tool, while the giants continue dominating the voice AI market with their closed APIs and massive infrastructures.
Sources: NVIDIA Research, TechCrunch, The Decoder
✓ Independent sources cross-checked and verified before publishing
