The Pulse
Hugging Face Transformers 5.18 Adds NVIDIA’s Nemotron 3 Diarization
Hugging Face released Transformers 5.18.0 on September 30, adding support for NVIDIA’s Nemotron 3 Diarization model. The open-weight model identifies up to eight speakers and supports both streaming and offline audio processing.
AI.info Team · Written with AI ·
Hugging Face released Transformers 5.18.0 on September 30 with built-in support for NVIDIA’s Nemotron 3 Diarization, an open-weight model that tracks who speaks when. The integration lets developers load the model through the Transformers library’s standard audio-classification interfaces. It supports live, chunk-by-chunk processing as well as offline analysis of recorded audio.
Eight speakers, one shared timeline
Nemotron 3 Diarization can identify as many as eight speakers, assigning speaker outputs according to the order each person first appears in the audio. That ordering is intended to keep speaker identities consistent as a conversation progresses. The model can process meetings, calls, podcasts and other conversational recordings, and can supply speaker labels and timestamps for speech-recognition systems.
The model uses an Arrival-Order Speaker Cache and a first-in, first-out queue to carry speaker information and recent audio context across streaming chunks. With chunked inference, the documentation says, the maximum recording duration is not limited. Output frame resolution can be configured in multiples of 10 milliseconds.
What the latency figures mean
The release notes describe configurable input-buffer settings from 80 milliseconds up to a 30.4-second offline-style buffer. Those figures describe audio buffering, not the time a user necessarily waits for a result. The Transformers documentation lists streaming modes with audio-wait latencies of 1.04 seconds, 0.64 seconds and 0.32 seconds, excluding computation time.
The model card says the buffer can be set as low as 80 milliseconds, but identifies 0.32 seconds as the lowest recommended configuration. That distinction matters for applications that need quick speaker updates: the smallest available buffer setting is not the same as a recommended end-to-end latency target.
Transformers makes the checkpoint easier to load
Developers can load the model with AutoProcessor and AutoModelForAudioFrameClassification, the general-purpose interfaces shown in the library’s examples. In streaming use, the processor splits audio into chunks, and the model returns a speaker cache that the next call can reuse. The final chunk is handled separately so the model can score remaining frames without look-ahead.
The NVIDIA model card lists the checkpoint at 99.2 million parameters and identifies its license as OpenMDW 1.1. The card says the model is intended for commercial or non-commercial use under that license. Transformers 5.18 adds library support; it does not introduce a new checkpoint or change the model’s license.
A library integration, not a new model launch
NVIDIA’s model card gives September 23, 2026, as the model’s release date, a week before the Transformers update. The September 30 release therefore marks its addition to Hugging Face’s library, making the existing checkpoint accessible through a familiar loading and inference path. The release notes also list other model additions and software changes, but Nemotron 3 Diarization is the update’s notable audio-processing addition.
For developers, the practical change is direct access to streaming diarization in Transformers without building a separate integration around the checkpoint. The model’s speaker limit, chunk handling and latency settings define what that support provides: speaker-attributed audio frames, with buffer choices that trade faster updates against more context.