fp16-safe conversions: segmentation fused-logsumexp tail, wespeaker guards to 0x1p-24

CHECKSUMS.sha256

@@ -0,0 +1,25 @@
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  ./CHECKSUMS.sha256
+f283c01fa863188733c33c6fddac4c5dca42ca7cb22580918e1bc55877897e69  ./pyannote_segmentation.mlmodelc/analytics/coremldata.bin
+e75fc0ac9641de87e3514369455c8c8a65b00aae339817b20642f115d5d8861e  ./pyannote_segmentation.mlmodelc/coremldata.bin
+2926b811344e40ab6ce5406354bf5aaac35a297d0e67e3c0d3f6dd766e9f5f8f  ./pyannote_segmentation.mlmodelc/metadata.json
+ded0d1ee11d77976b5c706ce667d0c8cb49977d3fe4367cccbd7b582bdb86dec  ./pyannote_segmentation.mlmodelc/model.mil
+0266f4ad4d843ecf31ef9220ad6b80616b3ec64a4404b64f3ea0371554e236ec  ./pyannote_segmentation.mlmodelc/weights/weight.bin
+45349a79cc3d2acdde0699d6c24a2d953e2f33600e35f2e3a3696a76ca0aa5d2  ./pyannote_segmentation.mlpackage/Data/com.apple.CoreML/model.mlmodel
+0266f4ad4d843ecf31ef9220ad6b80616b3ec64a4404b64f3ea0371554e236ec  ./pyannote_segmentation.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+d598581336168c03f0e782851d05f5f8405d5bdaaa7ea662ec4b9b4445a66f40  ./pyannote_segmentation.mlpackage/Manifest.json
+4d7649edc0b071fecf4778bf533de05d6eadd5ade828179d30beaadb31258798  ./wespeaker_int8.mlmodelc/analytics/coremldata.bin
+119a9f64487e7a6617b26b1c4420c48f4cf60397694ea24c6c3c53f6eaf88c3d  ./wespeaker_int8.mlmodelc/coremldata.bin
+b96c1cdd968230b5a79b75749d2f585779e2ca59ab8249f72a18a44e93471d0e  ./wespeaker_int8.mlmodelc/metadata.json
+7668b5eb56fbb3ba81c36a1a04aeadf0faa5f3d0d59ede28975909094ad22de0  ./wespeaker_int8.mlmodelc/model.mil
+fb7310f849dd3db544fb016b0073d773e5293a26763a980a06d0834526f3a848  ./wespeaker_int8.mlmodelc/weights/weight.bin
+5f8f14e1a044267294b1f554100cf9e3784e477ef575297b66f73d6251cc4317  ./wespeaker_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel
+fb7310f849dd3db544fb016b0073d773e5293a26763a980a06d0834526f3a848  ./wespeaker_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+d0435814f5e227f1f07624718e2d151ab7f81da1b0750331549bc5605f322c39  ./wespeaker_int8.mlpackage/Manifest.json
+30528b97b29e0f0221b99c1c3456484f3123a68214b777f324a8b85ef5634c9a  ./wespeaker.mlmodelc/analytics/coremldata.bin
+4a2840e7abc9aafa02ca23f6a3cb37fd8c8d9a95887336dae3d1e09f6ba7f9f6  ./wespeaker.mlmodelc/coremldata.bin
+330d8018e32a2c056ace110b3079aafacaaef37f70eae6cdad7296e85e9687c1  ./wespeaker.mlmodelc/metadata.json
+cff0cfe914078e9336754a9b38a68c2cdd88ca7b6bf97568ad551ab03ae1b666  ./wespeaker.mlmodelc/model.mil
+680837ec172d67c3197bba93800e1623eebfd35c3b17011802f5f98b8026a0aa  ./wespeaker.mlmodelc/weights/weight.bin
+d6c4bb8b77eec2118f81506cdc18c75923f1ba6e6961d1ce51b0eeb73e3a592c  ./wespeaker.mlpackage/Data/com.apple.CoreML/model.mlmodel
+680837ec172d67c3197bba93800e1623eebfd35c3b17011802f5f98b8026a0aa  ./wespeaker.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+a16e2bae11d61c54e2a7c400bfee151e1244233064f449ee28caf7ab6252b933  ./wespeaker.mlpackage/Manifest.json

README.md

@@ -0,0 +1,68 @@
+---
+license: other
+license_name: mixed-upstream
+license_link: https://github.com/pyannote/pyannote-audio
+tags:
+  - coreml
+  - speaker-diarization
+  - pyannote
+  - wespeaker
+  - apple-neural-engine
+library_name: coremlit
+---
+
+# 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](https://github.com/findit-studio/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](https://huggingface.co/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.

pyannote_segmentation.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f283c01fa863188733c33c6fddac4c5dca42ca7cb22580918e1bc55877897e69
+size 243

pyannote_segmentation.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e75fc0ac9641de87e3514369455c8c8a65b00aae339817b20642f115d5d8861e
+size 518

pyannote_segmentation.mlmodelc/metadata.json

@@ -0,0 +1,73 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "pyannote\/segmentation-3.0, epsilon-safe (reduce_log_sum_exp->sub tail)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 589 × 7)",
+        "shortDescription" : "",
+        "shape" : "[1, 589, 7]",
+        "name" : "segments",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float16, Float32)",
+    "modelParameters" : [
+
+    ],
+    "author" : "Phase-A reconversion: segmentation-3.0.onnx, stable fused log-softmax",
+    "specificationVersion" : 6,
+    "mlProgramOperationTypeHistogram" : {
+      "Abs" : 1,
+      "InstanceNorm" : 4,
+      "Cast" : 4,
+      "Conv" : 3,
+      "MaxPool" : 3,
+      "Lstm" : 4,
+      "Transpose" : 2,
+      "Linear" : 3,
+      "ReduceLogSumExp" : 1,
+      "Sub" : 1,
+      "LeakyRelu" : 5
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "12.0",
+      "tvOS" : "15.0",
+      "visionOS" : "1.0",
+      "watchOS" : "8.0",
+      "iOS" : "15.0",
+      "macCatalyst" : "15.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 160000)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 160000]",
+        "name" : "audio",
+        "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" : "pyannote_segmentation",
+    "method" : "predict"
+  }
+]

pyannote_segmentation.mlmodelc/model.mil

@@ -0,0 +1,148 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios15>(tensor<fp32, [1, 1, 160000]> audio) {
+            tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1]> wav_norm1d_weight_to_fp16 = const()[name = tensor<string, []>("wav_norm1d_weight_to_fp16"), val = tensor<fp16, [1]>([0x1.44p-7])];
+            tensor<fp16, [1]> wav_norm1d_bias_to_fp16 = const()[name = tensor<string, []>("wav_norm1d_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.734p-5])];
+            tensor<fp16, []> var_20_to_fp16 = const()[name = tensor<string, []>("op_20_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 160000]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1, 160000]> input_1_cast_fp16 = instance_norm(beta = wav_norm1d_bias_to_fp16, epsilon = var_20_to_fp16, gamma = wav_norm1d_weight_to_fp16, x = audio_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<string, []> var_38_pad_type_0 = const()[name = tensor<string, []>("op_38_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_38_strides_0 = const()[name = tensor<string, []>("op_38_strides_0"), val = tensor<int32, [1]>([10])];
+            tensor<int32, [2]> var_38_pad_0 = const()[name = tensor<string, []>("op_38_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_38_dilations_0 = const()[name = tensor<string, []>("op_38_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_38_groups_0 = const()[name = tensor<string, []>("op_38_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [80, 1, 251]> conv0_weight_to_fp16 = const()[name = tensor<string, []>("conv0_weight_to_fp16"), val = tensor<fp16, [80, 1, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 80, 15975]> var_38_cast_fp16 = conv(dilations = var_38_dilations_0, groups = var_38_groups_0, pad = var_38_pad_0, pad_type = var_38_pad_type_0, strides = var_38_strides_0, weight = conv0_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<fp16, [1, 80, 15975]> input_3_cast_fp16 = abs(x = var_38_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> var_44 = const()[name = tensor<string, []>("op_44"), val = tensor<int32, [1]>([3])];
+            tensor<int32, [1]> var_45 = const()[name = tensor<string, []>("op_45"), val = tensor<int32, [1]>([3])];
+            tensor<string, []> input_5_pad_type_0 = const()[name = tensor<string, []>("input_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<bool, []> input_5_ceil_mode_0 = const()[name = tensor<string, []>("input_5_ceil_mode_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 80, 5325]> input_5_cast_fp16 = max_pool(ceil_mode = input_5_ceil_mode_0, kernel_sizes = var_44, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = var_45, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [80]> norm0_weight_to_fp16 = const()[name = tensor<string, []>("norm0_weight_to_fp16"), val = tensor<fp16, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40320)))];
+            tensor<fp16, [80]> norm0_bias_to_fp16 = const()[name = tensor<string, []>("norm0_bias_to_fp16"), val = tensor<fp16, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40576)))];
+            tensor<fp16, []> var_49_to_fp16 = const()[name = tensor<string, []>("op_49_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 80, 5325]> input_7_cast_fp16 = instance_norm(beta = norm0_bias_to_fp16, epsilon = var_49_to_fp16, gamma = norm0_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, []> var_56_to_fp16 = const()[name = tensor<string, []>("op_56_to_fp16"), val = tensor<fp16, []>(0x1.47cp-7)];
+            tensor<fp16, [1, 80, 5325]> input_9_cast_fp16 = leaky_relu(alpha = var_56_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [60, 80, 5]> conv1_weight_to_fp16 = const()[name = tensor<string, []>("conv1_weight_to_fp16"), val = tensor<fp16, [60, 80, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40832)))];
+            tensor<fp16, [60]> conv1_bias_to_fp16 = const()[name = tensor<string, []>("conv1_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88896)))];
+            tensor<fp16, [1, 60, 5321]> input_11_cast_fp16 = conv(bias = conv1_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = conv1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<int32, [1]> var_73 = const()[name = tensor<string, []>("op_73"), val = tensor<int32, [1]>([3])];
+            tensor<int32, [1]> var_74 = const()[name = tensor<string, []>("op_74"), val = tensor<int32, [1]>([3])];
+            tensor<string, []> input_13_pad_type_0 = const()[name = tensor<string, []>("input_13_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_13_pad_0 = const()[name = tensor<string, []>("input_13_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<bool, []> input_13_ceil_mode_0 = const()[name = tensor<string, []>("input_13_ceil_mode_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 60, 1773]> input_13_cast_fp16 = max_pool(ceil_mode = input_13_ceil_mode_0, kernel_sizes = var_73, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = var_74, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [60]> norm1_weight_to_fp16 = const()[name = tensor<string, []>("norm1_weight_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89088)))];
+            tensor<fp16, [60]> norm1_bias_to_fp16 = const()[name = tensor<string, []>("norm1_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89280)))];
+            tensor<fp16, []> var_78_to_fp16 = const()[name = tensor<string, []>("op_78_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 60, 1773]> input_15_cast_fp16 = instance_norm(beta = norm1_bias_to_fp16, epsilon = var_78_to_fp16, gamma = norm1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, []> var_85_to_fp16 = const()[name = tensor<string, []>("op_85_to_fp16"), val = tensor<fp16, []>(0x1.47cp-7)];
+            tensor<fp16, [1, 60, 1773]> input_17_cast_fp16 = leaky_relu(alpha = var_85_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [60, 60, 5]> conv2_weight_to_fp16 = const()[name = tensor<string, []>("conv2_weight_to_fp16"), val = tensor<fp16, [60, 60, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89472)))];
+            tensor<fp16, [60]> conv2_bias_to_fp16 = const()[name = tensor<string, []>("conv2_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125568)))];
+            tensor<fp16, [1, 60, 1769]> input_19_cast_fp16 = conv(bias = conv2_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = conv2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<int32, [1]> var_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [1]>([3])];
+            tensor<int32, [1]> var_103 = const()[name = tensor<string, []>("op_103"), val = tensor<int32, [1]>([3])];
+            tensor<string, []> input_21_pad_type_0 = const()[name = tensor<string, []>("input_21_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_21_pad_0 = const()[name = tensor<string, []>("input_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<bool, []> input_21_ceil_mode_0 = const()[name = tensor<string, []>("input_21_ceil_mode_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 60, 589]> input_21_cast_fp16 = max_pool(ceil_mode = input_21_ceil_mode_0, kernel_sizes = var_102, pad = input_21_pad_0, pad_type = input_21_pad_type_0, strides = var_103, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [60]> norm2_weight_to_fp16 = const()[name = tensor<string, []>("norm2_weight_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125760)))];
+            tensor<fp16, [60]> norm2_bias_to_fp16 = const()[name = tensor<string, []>("norm2_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125952)))];
+            tensor<fp16, []> var_107_to_fp16 = const()[name = tensor<string, []>("op_107_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 60, 589]> input_23_cast_fp16 = instance_norm(beta = norm2_bias_to_fp16, epsilon = var_107_to_fp16, gamma = norm2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, []> var_114_to_fp16 = const()[name = tensor<string, []>("op_114_to_fp16"), val = tensor<fp16, []>(0x1.47cp-7)];
+            tensor<fp16, [1, 60, 589]> x_3_cast_fp16 = leaky_relu(alpha = var_114_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([2, 0, 1])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [512]> add_0 = const()[name = tensor<string, []>("add_0"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126144)))];
+            tensor<fp32, [512]> add_1 = const()[name = tensor<string, []>("add_1"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(128256)))];
+            tensor<fp32, [512, 60]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130368)))];
+            tensor<fp32, [512, 128]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253312)))];
+            tensor<fp32, [512, 60]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<fp32, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(515520)))];
+            tensor<fp32, [512, 128]> concat_7 = const()[name = tensor<string, []>("concat_7"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(638464)))];
+            tensor<fp32, [1, 256]> input_27_batch_first_lstm_h0_reshaped = const()[name = tensor<string, []>("input_27_batch_first_lstm_h0_reshaped"), val = tensor<fp32, [1, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(900672)))];
+            tensor<string, []> input_27_batch_first_direction_0 = const()[name = tensor<string, []>("input_27_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
+            tensor<bool, []> input_27_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_27_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_27_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_27_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_27_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_27_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_27_batch_first_activation_0 = const()[name = tensor<string, []>("input_27_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [589, 1, 60]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [589, 1, 60]> transpose_0_cast_fp16_to_fp32 = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [589, 1, 256]> input_27_batch_first_0, tensor<fp32, [1, 256]> input_27_batch_first_1, tensor<fp32, [1, 256]> input_27_batch_first_2 = lstm(activation = input_27_batch_first_activation_0, bias = add_0, bias_back = add_1, cell_activation = input_27_batch_first_cell_activation_0, direction = input_27_batch_first_direction_0, initial_c = input_27_batch_first_lstm_h0_reshaped, initial_h = input_27_batch_first_lstm_h0_reshaped, output_sequence = input_27_batch_first_output_sequence_0, recurrent_activation = input_27_batch_first_recurrent_activation_0, weight_hh = concat_5, weight_hh_back = concat_7, weight_ih = concat_4, weight_ih_back = concat_6, x = transpose_0_cast_fp16_to_fp32)[name = tensor<string, []>("input_27_batch_first")];
+            tensor<fp32, [512]> add_2 = const()[name = tensor<string, []>("add_2"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(901760)))];
+            tensor<fp32, [512]> add_3 = const()[name = tensor<string, []>("add_3"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(903872)))];
+            tensor<fp32, [512, 256]> concat_14 = const()[name = tensor<string, []>("concat_14"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(905984)))];
+            tensor<fp32, [512, 128]> concat_15 = const()[name = tensor<string, []>("concat_15"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1430336)))];
+            tensor<fp32, [512, 256]> concat_16 = const()[name = tensor<string, []>("concat_16"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1692544)))];
+            tensor<fp32, [512, 128]> concat_17 = const()[name = tensor<string, []>("concat_17"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2216896)))];
+            tensor<string, []> input_29_batch_first_direction_0 = const()[name = tensor<string, []>("input_29_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
+            tensor<bool, []> input_29_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_29_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_29_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_29_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_29_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_29_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_29_batch_first_activation_0 = const()[name = tensor<string, []>("input_29_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [589, 1, 256]> input_29_batch_first_0, tensor<fp32, [1, 256]> input_29_batch_first_1, tensor<fp32, [1, 256]> input_29_batch_first_2 = lstm(activation = input_29_batch_first_activation_0, bias = add_2, bias_back = add_3, cell_activation = input_29_batch_first_cell_activation_0, direction = input_29_batch_first_direction_0, initial_c = input_27_batch_first_lstm_h0_reshaped, initial_h = input_27_batch_first_lstm_h0_reshaped, output_sequence = input_29_batch_first_output_sequence_0, recurrent_activation = input_29_batch_first_recurrent_activation_0, weight_hh = concat_15, weight_hh_back = concat_17, weight_ih = concat_14, weight_ih_back = concat_16, x = input_27_batch_first_0)[name = tensor<string, []>("input_29_batch_first")];
+            tensor<fp32, [512]> add_4 = const()[name = tensor<string, []>("add_4"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2479104)))];
+            tensor<fp32, [512]> add_5 = const()[name = tensor<string, []>("add_5"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2481216)))];
+            tensor<fp32, [512, 256]> concat_24 = const()[name = tensor<string, []>("concat_24"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2483328)))];
+            tensor<fp32, [512, 128]> concat_25 = const()[name = tensor<string, []>("concat_25"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3007680)))];
+            tensor<fp32, [512, 256]> concat_26 = const()[name = tensor<string, []>("concat_26"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3269888)))];
+            tensor<fp32, [512, 128]> concat_27 = const()[name = tensor<string, []>("concat_27"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3794240)))];
+            tensor<string, []> input_31_batch_first_direction_0 = const()[name = tensor<string, []>("input_31_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
+            tensor<bool, []> input_31_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_31_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_31_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_31_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_31_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_31_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_31_batch_first_activation_0 = const()[name = tensor<string, []>("input_31_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [589, 1, 256]> input_31_batch_first_0, tensor<fp32, [1, 256]> input_31_batch_first_1, tensor<fp32, [1, 256]> input_31_batch_first_2 = lstm(activation = input_31_batch_first_activation_0, bias = add_4, bias_back = add_5, cell_activation = input_31_batch_first_cell_activation_0, direction = input_31_batch_first_direction_0, initial_c = input_27_batch_first_lstm_h0_reshaped, initial_h = input_27_batch_first_lstm_h0_reshaped, output_sequence = input_31_batch_first_output_sequence_0, recurrent_activation = input_31_batch_first_recurrent_activation_0, weight_hh = concat_25, weight_hh_back = concat_27, weight_ih = concat_24, weight_ih_back = concat_26, x = input_29_batch_first_0)[name = tensor<string, []>("input_31_batch_first")];
+            tensor<fp32, [512]> add_6 = const()[name = tensor<string, []>("add_6"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4056448)))];
+            tensor<fp32, [512]> add_7 = const()[name = tensor<string, []>("add_7"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4058560)))];
+            tensor<fp32, [512, 256]> concat_34 = const()[name = tensor<string, []>("concat_34"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4060672)))];
+            tensor<fp32, [512, 128]> concat_35 = const()[name = tensor<string, []>("concat_35"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4585024)))];
+            tensor<fp32, [512, 256]> concat_36 = const()[name = tensor<string, []>("concat_36"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4847232)))];
+            tensor<fp32, [512, 128]> concat_37 = const()[name = tensor<string, []>("concat_37"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5371584)))];
+            tensor<string, []> input_33_batch_first_direction_0 = const()[name = tensor<string, []>("input_33_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
+            tensor<bool, []> input_33_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_33_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_33_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_33_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_33_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_33_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_33_batch_first_activation_0 = const()[name = tensor<string, []>("input_33_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [589, 1, 256]> input_33_batch_first_0, tensor<fp32, [1, 256]> input_33_batch_first_1, tensor<fp32, [1, 256]> input_33_batch_first_2 = lstm(activation = input_33_batch_first_activation_0, bias = add_6, bias_back = add_7, cell_activation = input_33_batch_first_cell_activation_0, direction = input_33_batch_first_direction_0, initial_c = input_27_batch_first_lstm_h0_reshaped, initial_h = input_27_batch_first_lstm_h0_reshaped, output_sequence = input_33_batch_first_output_sequence_0, recurrent_activation = input_33_batch_first_recurrent_activation_0, weight_hh = concat_35, weight_hh_back = concat_37, weight_ih = concat_34, weight_ih_back = concat_36, x = input_31_batch_first_0)[name = tensor<string, []>("input_33_batch_first")];
+            tensor<int32, [3]> input_33_perm_0 = const()[name = tensor<string, []>("input_33_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<string, []> input_33_batch_first_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_33_batch_first_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [128, 256]> linear0_weight_to_fp16 = const()[name = tensor<string, []>("linear0_weight_to_fp16"), val = tensor<fp16, [128, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5633792)))];
+            tensor<fp16, [128]> linear0_bias_to_fp16 = const()[name = tensor<string, []>("linear0_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5699392)))];
+            tensor<fp16, [589, 1, 256]> input_33_batch_first_0_to_fp16 = cast(dtype = input_33_batch_first_0_to_fp16_dtype_0, x = input_33_batch_first_0)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 589, 256]> input_33_cast_fp16 = transpose(perm = input_33_perm_0, x = input_33_batch_first_0_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 589, 128]> linear_0_cast_fp16 = linear(bias = linear0_bias_to_fp16, weight = linear0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, []> var_234_to_fp16 = const()[name = tensor<string, []>("op_234_to_fp16"), val = tensor<fp16, []>(0x1.47cp-7)];
+            tensor<fp16, [1, 589, 128]> input_37_cast_fp16 = leaky_relu(alpha = var_234_to_fp16, x = linear_0_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<fp16, [128, 128]> linear1_weight_to_fp16 = const()[name = tensor<string, []>("linear1_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5699712)))];
+            tensor<fp16, [128]> linear1_bias_to_fp16 = const()[name = tensor<string, []>("linear1_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5732544)))];
+            tensor<fp16, [1, 589, 128]> linear_1_cast_fp16 = linear(bias = linear1_bias_to_fp16, weight = linear1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, []> var_239_to_fp16 = const()[name = tensor<string, []>("op_239_to_fp16"), val = tensor<fp16, []>(0x1.47cp-7)];
+            tensor<fp16, [1, 589, 128]> input_cast_fp16 = leaky_relu(alpha = var_239_to_fp16, x = linear_1_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [7, 128]> classifier_weight_to_fp16 = const()[name = tensor<string, []>("classifier_weight_to_fp16"), val = tensor<fp16, [7, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5732864)))];
+            tensor<fp16, [7]> classifier_bias_to_fp16 = const()[name = tensor<string, []>("classifier_bias_to_fp16"), val = tensor<fp16, [7]>([-0x1.01p+0, 0x1.67cp-2, 0x1.3d8p-1, 0x1.c8cp-2, -0x1.444p-2, -0x1.59p-1, -0x1.8fcp-2])];
+            tensor<fp16, [1, 589, 7]> linear_2_cast_fp16 = linear(bias = classifier_bias_to_fp16, weight = classifier_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [1]> var_247_axes_0 = const()[name = tensor<string, []>("op_247_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_247_keep_dims_0 = const()[name = tensor<string, []>("op_247_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 589, 1]> var_247_cast_fp16 = reduce_log_sum_exp(axes = var_247_axes_0, keep_dims = var_247_keep_dims_0, x = linear_2_cast_fp16)[name = tensor<string, []>("op_247_cast_fp16")];
+            tensor<fp16, [1, 589, 7]> var_249_cast_fp16 = sub(x = linear_2_cast_fp16, y = var_247_cast_fp16)[name = tensor<string, []>("op_249_cast_fp16")];
+            tensor<string, []> var_249_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_249_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 589, 7]> segments = cast(dtype = var_249_cast_fp16_to_fp32_dtype_0, x = var_249_cast_fp16)[name = tensor<string, []>("cast_0")];
+        } -> (segments);
+}

pyannote_segmentation.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0266f4ad4d843ecf31ef9220ad6b80616b3ec64a4404b64f3ea0371554e236ec
+size 5734720

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45349a79cc3d2acdde0699d6c24a2d953e2f33600e35f2e3a3696a76ca0aa5d2
+size 24289

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0266f4ad4d843ecf31ef9220ad6b80616b3ec64a4404b64f3ea0371554e236ec
+size 5734720

pyannote_segmentation.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "2AB5E6CE-0992-4457-BD72-544A89B64C74": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F4B9A266-717D-4897-82E1-28470E6A7FFC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F4B9A266-717D-4897-82E1-28470E6A7FFC"
+}

wespeaker.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30528b97b29e0f0221b99c1c3456484f3123a68214b777f324a8b85ef5634c9a
+size 243

wespeaker.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a2840e7abc9aafa02ca23f6a3cb37fd8c8d9a95887336dae3d1e09f6ba7f9f6
+size 512

wespeaker.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "WeSpeaker ResNet34-LM, epsilon-safe attentive-stat pooling divisor guard",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "constant",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 256)",
+        "shortDescription" : "",
+        "shape" : "[3, 256]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Phase-A reconversion: wespeaker ResNet34-LM, attentive-pool eps 1e-8 -> 2^-24",
+    "specificationVersion" : 6,
+    "mlProgramOperationTypeHistogram" : {
+      "Transpose" : 4,
+      "Square" : 2,
+      "UpsampleNearestNeighbor" : 1,
+      "Squeeze" : 5,
+      "Sub" : 5,
+      "ReduceMean" : 2,
+      "Gather" : 2,
+      "Identity" : 1,
+      "Reshape" : 1,
+      "Matmul" : 2,
+      "Concat" : 1,
+      "Add" : 19,
+      "Sqrt" : 1,
+      "RealDiv" : 3,
+      "Pad" : 2,
+      "Linear" : 2,
+      "Relu" : 33,
+      "ExpandDims" : 11,
+      "Conv" : 36,
+      "Maximum" : 1,
+      "Log" : 1,
+      "SliceByIndex" : 1001,
+      "Stack" : 1,
+      "ReduceSum" : 4,
+      "Mul" : 8
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "12.0",
+      "tvOS" : "15.0",
+      "visionOS" : "1.0",
+      "watchOS" : "8.0",
+      "iOS" : "15.0",
+      "macCatalyst" : "15.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 160000)",
+        "shortDescription" : "",
+        "shape" : "[3, 160000]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 589)",
+        "shortDescription" : "",
+        "shape" : "[3, 589]",
+        "name" : "mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.6.0"
+    },
+    "generatedClassName" : "wespeaker",
+    "method" : "predict"
+  }
+]

wespeaker.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:680837ec172d67c3197bba93800e1623eebfd35c3b17011802f5f98b8026a0aa
+size 28706752

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6c4bb8b77eec2118f81506cdc18c75923f1ba6e6961d1ce51b0eeb73e3a592c
+size 612953

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:680837ec172d67c3197bba93800e1623eebfd35c3b17011802f5f98b8026a0aa
+size 28706752

wespeaker.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "5D66401A-A11D-4CD9-88E0-39F3AAE8CBF8": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "DFACC015-BD66-478C-97BF-63A2559637EA": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "5D66401A-A11D-4CD9-88E0-39F3AAE8CBF8"
+}

wespeaker_int8.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d7649edc0b071fecf4778bf533de05d6eadd5ade828179d30beaadb31258798
+size 243

wespeaker_int8.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:119a9f64487e7a6617b26b1c4420c48f4cf60397694ea24c6c3c53f6eaf88c3d
+size 529

wespeaker_int8.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "WeSpeaker ResNet34-LM int8-palettized, epsilon-safe pooling guard",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "constant",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 256)",
+        "shortDescription" : "",
+        "shape" : "[3, 256]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float32, Palettized (8 bits))",
+    "modelParameters" : [
+
+    ],
+    "author" : "Phase-A reconversion: wespeaker ResNet34-LM, attentive-pool eps 1e-8 -> 2^-24 (int8 8-bit palettized)",
+    "specificationVersion" : 7,
+    "mlProgramOperationTypeHistogram" : {
+      "Transpose" : 4,
+      "UpsampleNearestNeighbor" : 1,
+      "Squeeze" : 5,
+      "Ios16.gather" : 2,
+      "Ios16.reduceMean" : 2,
+      "Ios16.linear" : 2,
+      "Ios16.add" : 19,
+      "Ios16.realDiv" : 3,
+      "Concat" : 1,
+      "Ios16.reduceSum" : 4,
+      "Ios16.square" : 2,
+      "Pad" : 2,
+      "Ios16.log" : 1,
+      "ExpandDims" : 11,
+      "Ios16.sub" : 5,
+      "Ios16.relu" : 33,
+      "Ios16.reshape" : 1,
+      "Ios16.constexprLutToDense" : 38,
+      "Ios16.matmul" : 2,
+      "SliceByIndex" : 1001,
+      "Stack" : 1,
+      "Ios16.maximum" : 1,
+      "Ios16.conv" : 36,
+      "Ios16.sqrt" : 1,
+      "Ios16.mul" : 8
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "13.0",
+      "tvOS" : "16.0",
+      "visionOS" : "1.0",
+      "watchOS" : "9.0",
+      "iOS" : "16.0",
+      "macCatalyst" : "16.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 160000)",
+        "shortDescription" : "",
+        "shape" : "[3, 160000]",
+        "name" : "waveform",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 3 × 589)",
+        "shortDescription" : "",
+        "shape" : "[3, 589]",
+        "name" : "mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.6.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "wespeaker_int8",
+    "method" : "predict"
+  }
+]

wespeaker_int8.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb7310f849dd3db544fb016b0073d773e5293a26763a980a06d0834526f3a848
+size 7243904

wespeaker_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f8f14e1a044267294b1f554100cf9e3784e477ef575297b66f73d6251cc4317
+size 618559

wespeaker_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb7310f849dd3db544fb016b0073d773e5293a26763a980a06d0834526f3a848
+size 7243904

wespeaker_int8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "37DC31EE-89C9-4957-B431-8BF5CFAA8EFF": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "BDEDC262-0C99-44CC-8757-C7E297D636DA": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "37DC31EE-89C9-4957-B431-8BF5CFAA8EFF"
+}