Add CamPlusPreprocessor_v2 + CamPlusPlus_v2: Kaldi-exact fbank, exact segment pooling, fp32 (FluidAudio #985)

FluidAudio PR https://github.com/FluidInference/FluidAudio/pull/1099 (issue https://github.com/FluidInference/FluidAudio/issues/985); recipe model-lab models/campplus-coreml-v2 (ef06123). Cosine 1.0000 to 3D-Speaker CAMPPlus at all lengths 0.35-60 s; AMI EER 2.85%/3.55% = PyTorch reference. v1 files are unchanged.

CamPlusPlus_v2.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0c5e0bf34974549ebc42431a49122d5cc05222a750a864ae53b9d9f625e0d5a
+size 243

CamPlusPlus_v2.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6412f16c728165dfbefce470359665ab94e3c0b7784d25febdd58faed035edab
+size 411

CamPlusPlus_v2.mlmodelc/metadata.json

@@ -0,0 +1,85 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "CAM++ speaker embedding (3D-Speaker campplus_cn_common), FluidAudio #985",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 192)",
+        "shortDescription" : "",
+        "shape" : "[1, 192]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.square" : 1,
+      "Ios17.reshape" : 53,
+      "Ios16.reduceMean" : 55,
+      "Ios17.transpose" : 1,
+      "Ios16.reduceProd" : 1,
+      "Ios17.expandDims" : 54,
+      "Ios16.avgPool" : 52,
+      "Ios17.add" : 56,
+      "Tile" : 52,
+      "Ios17.sliceByIndex" : 53,
+      "Ios16.sigmoid" : 52,
+      "Ios17.squeeze" : 1,
+      "Shape" : 157,
+      "Ios17.gather" : 104,
+      "Ios17.batchNorm" : 55,
+      "Ios17.conv" : 225,
+      "Ios17.sub" : 3,
+      "Ios16.relu" : 171,
+      "Ios17.cast" : 1,
+      "Ios17.realDiv" : 53,
+      "Ios17.concat" : 158,
+      "Ios17.sqrt" : 1,
+      "Ios17.mul" : 54
+    },
+    "computePrecision" : "Mixed (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" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 32...5998 × 80",
+        "shapeRange" : "[[1, 1], [32, 5998], [80, 80]]",
+        "formattedType" : "MultiArray (Float32 1 × 400 × 80)",
+        "type" : "MultiArray",
+        "shape" : "[1, 400, 80]",
+        "name" : "feats",
+        "shortDescription" : ""
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "generatedClassName" : "CamPlusPlus_v2",
+    "method" : "predict"
+  }
+]

CamPlusPlus_v2.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e5bafbc3d3faa94aff8c72afd22ffae97ce31d87ce70c000585e055cd776c18
+size 27670272

CamPlusPreprocessor_v2.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79a870e41d4cb1681d8a6d971f6df7958c06b7532ff5dfe7a6d554a5c18a775d
+size 243

CamPlusPreprocessor_v2.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a455197a14ef8f53a4ef8bc10a0b255660c124618b0c256b8b158896c3bb0e3d
+size 384

CamPlusPreprocessor_v2.mlmodelc/metadata.json

@@ -0,0 +1,71 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "CAM++ Kaldi fbank (Povey, dither 0), FluidAudio #985",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "features",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.clip" : 1,
+      "Ios17.conv" : 1,
+      "Ios17.log" : 1,
+      "Ios17.mul" : 2,
+      "Ios17.sliceByIndex" : 2,
+      "Ios17.matmul" : 1,
+      "Ios17.add" : 1,
+      "Ios17.transpose" : 1,
+      "Ios17.expandDims" : 1
+    },
+    "computePrecision" : "Mixed (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" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 400...960000",
+        "shapeRange" : "[[1, 1], [400, 960000]]",
+        "formattedType" : "MultiArray (Float32 1 × 64000)",
+        "type" : "MultiArray",
+        "shape" : "[1, 64000]",
+        "name" : "waveform",
+        "shortDescription" : ""
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.1"
+    },
+    "generatedClassName" : "CamPlusPreprocessor_v2",
+    "method" : "predict"
+  }
+]

CamPlusPreprocessor_v2.mlmodelc/model.mil

@@ -0,0 +1,37 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, ?]> waveform) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"waveform", [1, 64000]}}), ("RangeDims", {{"waveform", [[1, 1], [400, 960000]]}})))] {
+            tensor<fp32, [80, 257]> mel = const()[name = tensor<string, []>("mel"), val = tensor<fp32, [80, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [514, 1, 400]> kernel = const()[name = tensor<string, []>("kernel"), val = tensor<fp32, [514, 1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82368)))];
+            tensor<int32, [1]> var_10_axes_0 = const()[name = tensor<string, []>("op_10_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> var_10 = expand_dims(axes = var_10_axes_0, x = waveform)[name = tensor<string, []>("op_10")];
+            tensor<string, []> spec_pad_type_0 = const()[name = tensor<string, []>("spec_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> spec_strides_0 = const()[name = tensor<string, []>("spec_strides_0"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [2]> spec_pad_0 = const()[name = tensor<string, []>("spec_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> spec_dilations_0 = const()[name = tensor<string, []>("spec_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> spec_groups_0 = const()[name = tensor<string, []>("spec_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 514, ?]> spec = conv(dilations = spec_dilations_0, groups = spec_groups_0, pad = spec_pad_0, pad_type = spec_pad_type_0, strides = spec_strides_0, weight = kernel, x = var_10)[name = tensor<string, []>("spec")];
+            tensor<int32, [3]> re_begin_0 = const()[name = tensor<string, []>("re_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> re_end_0 = const()[name = tensor<string, []>("re_end_0"), val = tensor<int32, [3]>([1, 257, 0])];
+            tensor<bool, [3]> re_end_mask_0 = const()[name = tensor<string, []>("re_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp32, [1, 257, ?]> re = slice_by_index(begin = re_begin_0, end = re_end_0, end_mask = re_end_mask_0, x = spec)[name = tensor<string, []>("re")];
+            tensor<int32, [3]> im_begin_0 = const()[name = tensor<string, []>("im_begin_0"), val = tensor<int32, [3]>([0, 257, 0])];
+            tensor<int32, [3]> im_end_0 = const()[name = tensor<string, []>("im_end_0"), val = tensor<int32, [3]>([1, 514, 0])];
+            tensor<bool, [3]> im_end_mask_0 = const()[name = tensor<string, []>("im_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 257, ?]> im = slice_by_index(begin = im_begin_0, end = im_end_0, end_mask = im_end_mask_0, x = spec)[name = tensor<string, []>("im")];
+            tensor<fp32, [1, 257, ?]> var_52 = mul(x = re, y = re)[name = tensor<string, []>("op_52")];
+            tensor<fp32, [1, 257, ?]> var_53 = mul(x = im, y = im)[name = tensor<string, []>("op_53")];
+            tensor<fp32, [1, 257, ?]> power = add(x = var_52, y = var_53)[name = tensor<string, []>("power")];
+            tensor<bool, []> mel_1_transpose_x_0 = const()[name = tensor<string, []>("mel_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> mel_1_transpose_y_0 = const()[name = tensor<string, []>("mel_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 80, ?]> mel_1 = matmul(transpose_x = mel_1_transpose_x_0, transpose_y = mel_1_transpose_y_0, x = mel, y = power)[name = tensor<string, []>("mel_1")];
+            tensor<fp32, []> var_57 = const()[name = tensor<string, []>("op_57"), val = tensor<fp32, []>(0x1p-23)];
+            tensor<fp32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<fp32, []>(0x1.fffffep+127)];
+            tensor<fp32, [1, 80, ?]> clip_0 = clip(alpha = var_57, beta = const_0, x = mel_1)[name = tensor<string, []>("clip_0")];
+            tensor<fp32, []> var_60_epsilon_0 = const()[name = tensor<string, []>("op_60_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [1, 80, ?]> var_60 = log(epsilon = var_60_epsilon_0, x = clip_0)[name = tensor<string, []>("op_60")];
+            tensor<int32, [3]> var_63_perm_0 = const()[name = tensor<string, []>("op_63_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, ?, 80]> features = transpose(perm = var_63_perm_0, x = var_60)[name = tensor<string, []>("transpose_0")];
+        } -> (features);
+}

CamPlusPreprocessor_v2.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24c77ee75e17819622b3a69eeaccd57c6f6f05b0f5747c807a6540f032d3f1fb
+size 904832