NVIDIA Releases Nemotron 3 Diarization for Multi-Speaker Tracking
The design choice to output only speaker-and-time labels, leaving transcription to a separate ASR, makes it a modular component developers can slot into existing pipelines rather than an all-in-one solution. It sits in the same direction as Alibaba Qwen's role-based transcription ASR released the day before, offering a specialised alternative. Engineers building meeting tools or voice agents who need local, low-footprint diarization can test it directly.
NVIDIA launched Nemotron 3 Diarization, a speaker diarization model that tracks who spoke at what time, supporting up to 8 speakers.
It handles both live streams and recorded audio and can process overlapping speech.The 100M-parameter model is small enough to run locally.
It keeps speaker labels consistent across time and audio chunks using first-appearance speaker numbering, an AOSC long-term speaker cache, and a FIFO short-term context.
It outputs only speaker and timing information, so text transcription requires pairing it with a separate ASR model.
英伟达最新上的多说话人日志模型：Nemotron 3 Diarization，用来跟踪谁在什么时间说话，支持8说话人
直播与录音都支持、能处理重叠语音；100M，可本地跑
跟昨天阿里Qwen的那个分角色转写ASR是同一方向，英伟达这个等于是这个方向的专精组件版，不同的是英伟达这个只出“谁+何时”，文本需另配ASR
它通过“按首次出现顺序固定speaker编号+AOSC长期说话人缓存+FIFO短期上下文”来维持跨时间/跨chunk的说话人标签一致性，来避免把同一个人误判成新speaker、把不同人混成同一个speaker
#语音日志模型 #Nemotron3Diarization
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Nemotron 3 Diarization, NVIDIA, speaker diarization, Qwen.
Source attribution: curated from AIGCLINK (@aigclink). Cross-source corroboration: 11 independent sources have covered this development (including NVIDIA AI, shao__meng, Hugging Face: Blog), confirming industry-wide signal strength.
