fp16-safe conversion: fused x-logsumexp(x) tail; encoder weights byte-identical to upstream pin

CHECKSUMS.sha256

@@ -0,0 +1,9 @@
+d07ca6d8c69f2921810ca9bc919f0334f78de40a8f0c50e125358765fc5769fd  ./base960h_aligner.mlmodelc/analytics/coremldata.bin
+64b5214b50e003435d6659b8a3e4178e58ba48f3738daca0d6a1101412cb5a55  ./base960h_aligner.mlmodelc/coremldata.bin
+b7e9bf36c9290d843804f4082b5442098fd14d26f2e0b032225ddebe3df729b2  ./base960h_aligner.mlmodelc/metadata.json
+2d55c09bd32ee0810214d3896d0a5e87b2e8643b77c034bfcc908836e10d3809  ./base960h_aligner.mlmodelc/model.mil
+de51193fe73fb3aad085f9c794f08bfde1b939fc12f92e0834edcd4cb712e642  ./base960h_aligner.mlmodelc/weights/weight.bin
+ff4b7c3781d2a07698cb8883050697a45abb9e21325da3d9254df35139553c9c  ./base960h_aligner.mlpackage/Data/com.apple.CoreML/model.mlmodel
+de51193fe73fb3aad085f9c794f08bfde1b939fc12f92e0834edcd4cb712e642  ./base960h_aligner.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+bae000742a392d10914bce5d02bcb86ba4bbda83de8c7a141d27c79035fab0d3  ./base960h_aligner.mlpackage/Manifest.json
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  ./CHECKSUMS.sha256

README.md

@@ -0,0 +1,60 @@
+---
+license: apache-2.0
+tags:
+  - coreml
+  - forced-alignment
+  - wav2vec2
+  - ctc
+  - apple-neural-engine
+library_name: coremlit
+---
+
+# 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](https://github.com/findit-studio/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](https://huggingface.co/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.

base960h_aligner.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d07ca6d8c69f2921810ca9bc919f0334f78de40a8f0c50e125358765fc5769fd
+size 243

base960h_aligner.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64b5214b50e003435d6659b8a3e4178e58ba48f3738daca0d6a1101412cb5a55
+size 534

base960h_aligner.mlmodelc/metadata.json

@@ -0,0 +1,81 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "wav2vec2-base-960h forced aligner, epsilon-safe (reduce_log_sum_exp->sub tail)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 2999 × 29)",
+        "shortDescription" : "",
+        "shape" : "[1, 2999, 29]",
+        "name" : "emissions",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Phase-A reconversion: wav2vec2-base-960h aligner, stable fused log-softmax",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 13,
+      "Ios17.sqrt" : 1,
+      "Ios17.square" : 1,
+      "Ios17.transpose" : 51,
+      "Ios17.sub" : 2,
+      "Ios17.linear" : 74,
+      "Ios17.conv" : 8,
+      "Ios17.matmul" : 24,
+      "Ios16.reduceLogSumExp" : 1,
+      "Ios17.add" : 27,
+      "Ios17.realDiv" : 1,
+      "Ios16.reduceMean" : 2,
+      "Ios17.sliceByIndex" : 1,
+      "Ios16.softmax" : 12,
+      "Ios16.gelu" : 20,
+      "Ios17.expandDims" : 1,
+      "Ios17.reshape" : 50,
+      "Ios17.cast" : 2,
+      "Ios17.layerNorm" : 26
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 960000)",
+        "shortDescription" : "",
+        "shape" : "[1, 960000]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-16",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "base960h_aligner",
+    "method" : "predict"
+  }
+]

base960h_aligner.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de51193fe73fb3aad085f9c794f08bfde1b939fc12f92e0834edcd4cb712e642
+size 188799802

base960h_aligner.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff4b7c3781d2a07698cb8883050697a45abb9e21325da3d9254df35139553c9c
+size 154521

base960h_aligner.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de51193fe73fb3aad085f9c794f08bfde1b939fc12f92e0834edcd4cb712e642
+size 188799802

base960h_aligner.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "4296390D-EF20-44EA-B09C-2E9B874A1BA0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "EC9FDB7A-F224-4BE0-BE4A-1BD1C62458C4": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "EC9FDB7A-F224-4BE0-BE4A-1BD1C62458C4"
+}