Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages (huggingface.co)

NVIDIA has announced the launch of Nemotron 3.5 Transcribe, an advanced open ASR model that supports multilingual transcription, speaker attribution, translation, and audio Q&A across 40 languages. This model, now available on Hugging Face and ModelScope, is a game-changer for applications requiring nuanced context, such as multilingual calls or medical dictation. It excels in transcribing recordings of up to an hour, managing speaker labels for up to eight individuals, and allowing developers to guide output through natural language prompts. This model has achieved top rankings on various ASR leaderboards, including the Hugging Face Open ASR Leaderboard, where it recorded an impressive 5.33% average word error rate (WER) for English, and showcases low error rates in speaker attribution across multiple benchmarks. The significance of Nemotron 3.5 Transcribe lies in its ability to integrate various complex tasks into one model without sacrificing accuracy, making it highly adaptable for diverse applications. With a focus on natural language prompts and domain-specific adaptations, the model can deliver tailored outputs for different contexts, enhancing the utility of speech technology in real-world scenarios. Complementing Nemotron 3.5 is the Granary v2 data pipeline, which supplies a robust framework for fine-tuning and customization, ultimately driving improved performance metrics, such as reducing word error rates significantly in specialized fields like healthcare. Together, these advancements solidify NVIDIA's position at the forefront of AI-driven speech processing technology.