NVIDIA Announces Nemotron 3 Diarization for Real-Time Multi-Speaker AI

NVIDIA has unveiled Nemotron 3 Diarization, an open-weight model capable of real-time speaker diarization with up to eight speakers, achieving a 14.72% Diarization Error Rate (DER) on VoiceArena's leaderboard. The model supports both offline and streaming conversations, handling overlapping speech and customizable streaming latency. Nemotron 3 Diarization is designed to enhance applications such as meeting transcripts, customer calls, and podcasts by accurately attributing speech to individual speakers. The model's capabilities are demonstrated in a demo with Argmax Pro SDK 3, showcasing real-time speaker attribution in complex conversations. This advancement in speaker diarization technology aims to improve the accuracy and utility of speech recognition systems in various real-world scenarios.

Source: huggingface