Diarize Speakers (pyannote)

How it fits

Typical backends

Caveats

• CompatibilityThe pyannote community-1 pipeline is bound at the model slot and loaded fully offline at startup — there is no runtime HuggingFace download and no access token in the component config. The repo is self-contained: its config references bundled segmentation, embedding, and VBx/PLDA clustering sub-models, fetched once when the file is linked; a slot missing the config or its sub-models aborts startup with a load error.
• I/O contractAn empty input AudioFrame returns an empty list without invoking the pipeline. Output start/end times are in seconds relative to the start of the input frame, not absolute stream time, and multi-channel input is averaged to mono before resampling.
• State lifetimeThe speaker label is an opaque per-call identifier and is not stable across frames or runs; pair with a downstream embedding step if persistent speaker identity is required.
• Parameter interactionmin_speakers and max_speakers are applied only when greater than zero; mis-set bounds silently over-merge or over-split turns.
• AccuracyAnything not at 16 kHz is resampled internally, and supplying audio captured below 16 kHz hurts accuracy noticeably.
• CostPer-call cost scales with input length and the clustering step is super-linear, so very long frames are expensive; device starting with `cuda` silently falls back to CPU when CUDA is unavailable.