Nvidia Nemotron 3.5 Transcribe (huggingface.co)

Nvidia has announced the release of Nemotron 3.5 Transcribe, a cutting-edge offline speech and audio model designed for transcription across 40 languages. This model features a hybrid mixture-of-experts architecture that leverages approximately 4.3 billion parameters and can handle recordings of up to one hour with an impressive 65,536-token context. It supports various tasks, including multilingual transcription, speech translation, and audio question answering, making it highly versatile for commercial and non-commercial uses. The significance of Nemotron 3.5 Transcribe lies in its high accuracy and flexible application across different speech processing tasks. With state-of-the-art performance benchmarks—such as a mean word error rate (WER) of 4.71 for English ASR—the model sets a new standard in the field. Notably, it employs advanced techniques like multi-token prediction and speculative decoding to optimize processing efficiency. This makes it well-suited for developers and researchers working in areas like customer service analytics, transcription tools, and audio-understanding applications. The model's release in multiple precisions (BF16, FP8, NVFP4) further enhances its accessibility for varied deployment scenarios.