bjnortier/coreai-ht5-nemotron-3-diarization-float16-split

Nemotron-3-Diarization — Core AI export (float16, split)

The copy Hello Transcribe 5.1 and later download, pinned by commit. It is the nemotron-3-diarization_float16_split export with its top-level folder renamed ht5-nemotron-3-diarization_float16_split; nothing inside is changed.

An Apple Core AI export of NVIDIA's Nemotron-3-Diarization, for on-device streaming speaker diarization (up to 8 speakers) on Apple silicon.

These are not PyTorch weights. The repository holds a single archive of .aimodel assets that run through Apple's Core AI runtime on macOS 27 / iOS 27 and later. They will not load with NeMo.

Contents

ht5-nemotron-3-diarization_float16_split.zip (~619 MB) unpacks to a bundle directory, ht5-nemotron-3-diarization_float16_split/:

File	Purpose
metadata.json	Bundle descriptor: features, speaker cache, and one entry per latency mode
step_offline.aimodel	Step graph for offline (whole-file) diarization
step_low_latency.aimodel	Step graph for low-latency streaming
step_very_low_latency.aimodel	Step graph for very-low-latency streaming
step_ultra_low_latency.aimodel	Step graph for ultra-low-latency streaming
silence_embeds.f32	Silence embeddings that seed the speaker cache
fixtures/	Reference tensors for checking a runtime's parity with the source model

The "split" export has one step graph per latency mode (modes.<mode>.asset in metadata.json), so a runtime loads only the modes it uses.

Configuration

• Precision: float16 · Speakers: 8
• Mel bins: 128 · Subsampling: 8× · Hidden size: 512
• Audio in: 16 kHz mono float32

License and attribution

Released under the OpenMDW 1.1 licence, the licence of the source model. The original Nemotron-3-Diarization is by NVIDIA; this repository only changes its serialization format. Refer to the upstream model card for training data, evaluation results, and intended use.