speakerkit-coreml — fp16-safe CoreML conversions for speaker diarization

CoreML conversions of the pyannote community-1 diarization pipeline's neural components, re-converted with fp16-survivable numerical guards so they are correct on the Apple Neural Engine and GPU — not just CPU.

Consumed by the speakerkit crate of the coremlit workspace (Rust, sync, sans-I/O), feeding the dia diarization pipeline's clustering.

Why these conversions exist

Stock conversions of these models carry numerical guard epsilons below fp16's smallest subnormal (2^-24). When a graph runs on the ANE (which computes in fp16 regardless of the declared dtype), those guards round to zero:

• the segmentation model's softmax → log(eps≈0) tail saturates log(0) to ≈ −45440, which in the downstream pipeline collapsed an 8-speaker recording to 5 detected speakers (16.6% DER, 100% confusion) while the ONNX reference was frame-perfect;
• the embedder's pooling real_div guards (1e-8) vanish the same way.

Even a naive re-export reproduces the defect: torch.log_softmax still lowers to the decomposed softmax → log(eps=0) under coremltools 9.0. These conversions use an explicitly fused x − logsumexp(x) tail and raise every guard constant to 0x1p-24.

Contents

artifact	form	what it is
pyannote_segmentation.{mlpackage,mlmodelc}	fp16	pyannote segmentation-3.0 (PyanNet), fused log-probability tail
wespeaker.{mlpackage,mlmodelc}	fp32 weights	WeSpeaker ResNet34-LM speaker embedder, pooling guards raised to 0x1p-24 (otherwise bit-identical numerics to the stock conversion)
wespeaker_int8.{mlpackage,mlmodelc}	8-bit palettized	same fixed embedder, k-means per-tensor palettization (~3.7× smaller, the production default)

.mlpackage is the canonical distributable; .mlmodelc is the compiled form the test suite consumes directly. CHECKSUMS.sha256 covers every file.

Verification (summary)

• Every graph passes a MIL-level fp16-guard audit (every guard's effective floor ≥ 2^-24, constants followed through cast, unresolvable guards fail the audit).
• Segmentation: worst per-window delta vs the ONNX oracle dropped from 45,422 → 0.07 (ANE), argmax agreement 100%; the 8-speaker collapse is fixed at model level.
• Embedder: All-vs-CPU cosine 1.0 on real speech; int8 clusters identically to fp32 on the multi-speaker parity corpus (speaker counts never differ).
• Conversion toolchain pinned in the recipe: coremltools 9.0, torch 2.5.1, onnx 1.22.0, python 3.11.15.

Upstream provenance & licensing

component	upstream	license
segmentation	pyannote/segmentation-3.0 (weights via its ONNX export)	MIT (upstream repo is access-gated; the MIT license permits redistribution with attribution — please also respect pyannote's usage conditions and cite their papers)
embedder	WeSpeaker ResNet34-LM (voxceleb)	Apache-2.0 (WeSpeaker project)

These are derivative conversions: same weights, corrected graph numerics. If you are the upstream author of either model and want changes to this redistribution, open a discussion.