AI News (2026/10/30): NVIDIA Open-Sources Nemotron 3.5 Transcribe Speech Model

News Details

NVIDIA open-sourced its speech transcription model Nemotron 3.5 Transcribe on GitHub on October 27. Based on a Mixture-of-Experts (MoE) architecture, this model demonstrates significant performance improvements in speech recognition and cross-language translation.

• [MoE Architecture Breakthrough]: The model contains 34 billion parameters, activating a 4.3 billion parameter expert subset during inference to balance efficiency and accuracy through dynamic computational resource allocation.
• [Multilingual Processing Capability]: Supports 40-language transcription with English mutual translation for 31 of these languages, and features real-time speaker diarization capable of distinguishing up to 8 speakers.
• [Optimized Deployment Costs]: Modular design reduces deployment barriers, enabling 1-hour audio processing tasks on an RTX 4090 GPU - approximately 60% less hardware requirement compared to traditional solutions.

---

AI-ALL In-Depth Analysis

The open-sourcing of Nemotron 3.5 Transcribe marks a new phase for scalable speech AI applications. Its MoE architecture achieves high-precision transcription (5.33% average word error rate on Open ASR English benchmarks) while maintaining low latency, directly challenging Meta's Whisper series in open-source dominance.

The model offers three key values to developers:

1. Simplified Cross-Language Engineering: Unified framework for handling 40 languages and translation needs
2. Comprehensive Multi-Speaker Scenarios: Real-time speaker diarization directly applicable to complex use cases like meeting recordings
3. Reinvented Compute Costs: Enables large model inference tasks on 4090-class GPUs

Notably, its 2.8% word error rate in non-streaming tasks already surpasses most commercial solutions. This open approach could accelerate speech technology adoption in vertical fields like education and healthcare.

---