alignkit-coreml — fp16-safe CoreML conversion of wav2vec2-base-960h for forced alignment

A CoreML conversion of facebook/wav2vec2-base-960h (CTC head, 29-class character vocabulary) with an fp16-survivable log-probability tail, for word-level forced alignment. Consumed by the alignkit crate of the coremlit workspace.

Why this conversion exists

The stock conversion's tail is a decomposed softmax → log(eps = 2^-149). That epsilon is below fp16's smallest subnormal (2^-24), so on the Apple Neural Engine the guard rounds to zero and log(0) saturates to ≈ −45440, corrupting 16.7% of emission cells and shifting word timings by up to 881 ms — silently, since the values are still finite and negative. (A naive torch.log_softmax re-export reproduces the defect under coremltools 9.0.) This conversion uses an explicitly fused x − logsumexp(x) tail: the emissions are exact log-probabilities on CPU and bounded (min > −100) on every placement.

Contents

• base960h_aligner.{mlpackage,mlmodelc} — encoder weights byte-identical to the established conversion (weight.bin sha256 de51193f…); only the tail graph changed.
• CHECKSUMS.sha256 — per-file digests.

Input: 16 kHz mono f32, up to a 960,000-sample (60 s) window. Output: [1, T, 29] log-probabilities, 320-sample hop.

Placement guidance (measured)

• CPU: exact log-probabilities (per-frame logsumexp ≈ 0) and the fastest placement at this model size — the recommended default.
• GPU / ANE: numerically safe with this conversion (emissions bounded, no saturation), but the wav2vec2 encoder itself is fp16-sensitive (group-norm): expect ~98% frame-argmax agreement vs fp32, i.e. small word-timing differences. Consumers that require exact timings should stay on CPU; the crate's runtime guard rejects any corrupted emission matrix regardless of placement.

Verification (summary)

MIL-level fp16-guard audit clean (27 guard sites, every effective floor ≥ 2^-24); emissions min > −100 on ANE (−23.81) and GPU (−28.36) with real speech; word-timing parity vs an independent ONNX implementation: median 0.0 ms on unpadded 60 s audio (367/372 boundaries within one 20 ms frame). Toolchain pinned: coremltools 9.0, torch/torchaudio 2.5.1, python 3.11.15.

Upstream provenance & licensing

Weights: facebook/wav2vec2-base-960h (Apache-2.0), via torchaudio's WAV2VEC2_ASR_BASE_960H pipeline. This is a derivative conversion: identical weights, corrected graph numerics.