Split FBank and Emebdding

Embedding.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1f5edc9b97ea81d5d526c4190bb546a83b6d3cb3c861a2caf8f296d6d7cc96c1
+oid sha256:ab79be8f578f5b54b085cb38190c91e01bf1e71ab9b1189a19fe009f70dbbf52
 size 243

Embedding.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3eedfaf32d775ddf6f06ae784b21177d7683ab647509b7ce4857548cbd112a73
-size 603
+oid sha256:ef4e9c40ab61e1f3c1f04e05133ff2cedb385d75207823c15fdcce82f8a296a1
+size 614

Embedding.mlmodelc/metadata.json

@@ -1,15 +1,15 @@
 [
   {
-    "shortDescription" : "pyannote community-1 speaker embedding (10 s WeSpeaker ResNet34, interpolates 589-frame weights to pooling layer size internally)",
+    "shortDescription" : "pyannote community-1 speaker embedding (10 s WeSpeaker ResNet34, FBANK + segmentation weights input)",
     "metadataOutputVersion" : "3.0",
     "outputSchema" : [
       {
         "hasShapeFlexibility" : "0",
         "isOptional" : "0",
         "dataType" : "Float32",
-        "formattedType" : "MultiArray (Float32 1 × 256)",
+        "formattedType" : "MultiArray (Float32)",
         "shortDescription" : "",
-        "shape" : "[1, 256]",
+        "shape" : "[]",
         "name" : "embedding",
         "type" : "MultiArray"
       }
@@ -20,33 +20,27 @@
     ],
     "author" : "Fluid Inference",
     "specificationVersion" : 8,
-    "storagePrecision" : "Float32",
+    "storagePrecision" : "Float16",
     "license" : "CC-BY-4.0",
     "mlProgramOperationTypeHistogram" : {
-      "Ios16.reduceL2Norm" : 1,
-      "Ios17.reshape" : 1,
-      "Ios16.reduceMean" : 2,
-      "Ios17.transpose" : 2,
-      "Ios17.expandDims" : 8,
-      "Ios17.add" : 20,
-      "Ios17.gatherAlongAxis" : 2,
-      "Ios17.sliceByIndex" : 1,
-      "Ios16.reduceSum" : 4,
-      "Ios17.squeeze" : 4,
-      "Pad" : 2,
-      "Ios17.log" : 1,
+      "Ios17.mul" : 4,
       "Ios17.sqrt" : 1,
-      "Ios17.sub" : 5,
+      "Ios17.sub" : 2,
       "Ios17.conv" : 37,
-      "Ios17.clip" : 3,
-      "Ios16.relu" : 33,
-      "Ios17.linear" : 4,
-      "Ios17.pow" : 2,
-      "Ios17.realDiv" : 4,
       "Ios17.concat" : 1,
-      "Ios17.mul" : 9
+      "Ios17.add" : 18,
+      "Ios17.realDiv" : 4,
+      "UpsampleNearestNeighbor" : 1,
+      "Ios16.relu" : 33,
+      "Ios17.clip" : 2,
+      "Ios17.expandDims" : 7,
+      "Ios16.reduceSum" : 4,
+      "Ios16.reduceL2Norm" : 1,
+      "Ios17.squeeze" : 2,
+      "Ios17.reshape" : 2,
+      "Ios17.cast" : 13
     },
-    "computePrecision" : "Mixed (Float32, Int32)",
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
     "stateSchema" : [
 
     ],
@@ -64,28 +58,32 @@
     },
     "inputSchema" : [
       {
-        "hasShapeFlexibility" : "0",
-        "isOptional" : "0",
         "dataType" : "Float32",
-        "formattedType" : "MultiArray (Float32 1 × 1 × 160000)",
-        "shortDescription" : "",
-        "shape" : "[1, 1, 160000]",
-        "name" : "audio",
-        "type" : "MultiArray"
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1...32 × 1 × 80 × 998",
+        "shapeRange" : "[[1, 32], [1, 1], [80, 80], [998, 998]]",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 80 × 998)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 80, 998]",
+        "name" : "fbank",
+        "shortDescription" : ""
       },
       {
-        "hasShapeFlexibility" : "0",
-        "isOptional" : "0",
         "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1...32 × 589",
+        "shapeRange" : "[[1, 32], [589, 589]]",
         "formattedType" : "MultiArray (Float32 1 × 589)",
-        "shortDescription" : "",
+        "type" : "MultiArray",
         "shape" : "[1, 589]",
         "name" : "weights",
-        "type" : "MultiArray"
+        "shortDescription" : ""
       }
     ],
     "userDefinedMetadata" : {
-      "com.github.apple.coremltools.conversion_date" : "2025-10-03",
+      "com.github.apple.coremltools.conversion_date" : "2025-10-13",
       "com.github.apple.coremltools.source" : "torch==2.8.0",
       "com.github.apple.coremltools.version" : "9.0b1",
       "com.github.apple.coremltools.source_dialect" : "TorchScript"

Embedding.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a771617a3a551b18874c78ae5f92f6f660a14e84e55c0121e6402725f88d6e2
-size 28304640
+oid sha256:6dba18a57a81b1e872802ca4def29541bb7900ccff430d9b2040092cadd7d688
+size 13264960

FBank.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a57eb96c61ea3470b3c509eafd8085e3a05d023219096414554700bc31e3f25
+size 243

FBank.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33b0f747fa880e3446f676078fd1b33f7333db5e94cb956b321c24f14b2c12bf
+size 538

FBank.mlmodelc/metadata.json

@@ -0,0 +1,81 @@
+[
+  {
+    "shortDescription" : "pyannote community-1 FBANK front-end (10 s audio -> 998 frames x 80 bins)",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "fbank",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "pyannote-speaker-diarization-community-1",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "storagePrecision" : "Float16",
+    "license" : "CC-BY-4.0",
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 2,
+      "Ios17.transpose" : 1,
+      "Ios17.sub" : 2,
+      "Ios17.conv" : 4,
+      "Ios17.log" : 1,
+      "Ios17.sliceByIndex" : 1,
+      "Ios16.reduceMean" : 1,
+      "Ios17.add" : 1,
+      "Ios17.clip" : 1,
+      "Ios17.pow" : 2,
+      "Ios17.expandDims" : 3,
+      "Ios17.squeeze" : 4,
+      "Ios17.reshape" : 2,
+      "Ios17.cast" : 6,
+      "Pad" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "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 × 1 × 160000",
+        "shapeRange" : "[[1, 32], [1, 1], [160000, 160000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 160000)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 160000]",
+        "name" : "audio",
+        "shortDescription" : ""
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-10-13",
+      "com.github.apple.coremltools.source" : "torch==2.8.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "fbank_community_1",
+    "method" : "predict"
+  }
+]

FBank.mlmodelc/model.mil

@@ -0,0 +1,97 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.8.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [?, 1, 160000]> audio) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1, 160000]}}), ("RangeDims", {{"audio", [[1, 32], [1, 1], [160000, 160000]]}})))] {
+            tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> frames_1_pad_type_0 = const()[name = tensor<string, []>("frames_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> frames_1_strides_0 = const()[name = tensor<string, []>("frames_1_strides_0"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [2]> frames_1_pad_0 = const()[name = tensor<string, []>("frames_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> frames_1_dilations_0 = const()[name = tensor<string, []>("frames_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> frames_1_groups_0 = const()[name = tensor<string, []>("frames_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [400, 1, 400]> frame_kernel_to_fp16 = const()[name = tensor<string, []>("frame_kernel_to_fp16"), val = tensor<fp16, [400, 1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [?, 1, 160000]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [?, 400, 998]> frames_1_cast_fp16 = conv(dilations = frames_1_dilations_0, groups = frames_1_groups_0, pad = frames_1_pad_0, pad_type = frames_1_pad_type_0, strides = frames_1_strides_0, weight = frame_kernel_to_fp16, x = audio_to_fp16)[name = tensor<string, []>("frames_1_cast_fp16")];
+            tensor<int32, [3]> frames_3_perm_0 = const()[name = tensor<string, []>("frames_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [2]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [2]>([-1, 400])];
+            tensor<fp16, [?, 998, 400]> frames_3_cast_fp16 = transpose(perm = frames_3_perm_0, x = frames_1_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [?, 400]> frames_5_cast_fp16 = reshape(shape = concat_0x, x = frames_3_cast_fp16)[name = tensor<string, []>("frames_5_cast_fp16")];
+            tensor<string, []> frames_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("frames_5_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_50_axes_0 = const()[name = tensor<string, []>("op_50_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_50_keep_dims_0 = const()[name = tensor<string, []>("op_50_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [?, 400]> frames_5_cast_fp16_to_fp32 = cast(dtype = frames_5_cast_fp16_to_fp32_dtype_0, x = frames_5_cast_fp16)[name = tensor<string, []>("cast_6")];
+            tensor<fp32, [?, 1]> var_50 = reduce_mean(axes = var_50_axes_0, keep_dims = var_50_keep_dims_0, x = frames_5_cast_fp16_to_fp32)[name = tensor<string, []>("op_50")];
+            tensor<string, []> var_50_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_50_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [?, 1]> var_50_to_fp16 = cast(dtype = var_50_to_fp16_dtype_0, x = var_50)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [?, 400]> frames_7_cast_fp16 = sub(x = frames_5_cast_fp16, y = var_50_to_fp16)[name = tensor<string, []>("frames_7_cast_fp16")];
+            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [?, 1, 400]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = frames_7_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [6]> var_60_pad_0 = const()[name = tensor<string, []>("op_60_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
+            tensor<string, []> var_60_mode_0 = const()[name = tensor<string, []>("op_60_mode_0"), val = tensor<string, []>("replicate")];
+            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [?, 1, 401]> var_60_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = var_60_mode_0, pad = var_60_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("op_60_cast_fp16")];
+            tensor<int32, [1]> padded_axes_0 = const()[name = tensor<string, []>("padded_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [?, 401]> padded_cast_fp16 = squeeze(axes = padded_axes_0, x = var_60_cast_fp16)[name = tensor<string, []>("padded_cast_fp16")];
+            tensor<int32, [2]> var_72_begin_0 = const()[name = tensor<string, []>("op_72_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_72_end_0 = const()[name = tensor<string, []>("op_72_end_0"), val = tensor<int32, [2]>([0, 400])];
+            tensor<bool, [2]> var_72_end_mask_0 = const()[name = tensor<string, []>("op_72_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [?, 400]> var_72_cast_fp16 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = padded_cast_fp16)[name = tensor<string, []>("op_72_cast_fp16")];
+            tensor<fp16, []> var_73_to_fp16 = const()[name = tensor<string, []>("op_73_to_fp16"), val = tensor<fp16, []>(0x1.f0cp-1)];
+            tensor<fp16, [?, 400]> var_74_cast_fp16 = mul(x = var_72_cast_fp16, y = var_73_to_fp16)[name = tensor<string, []>("op_74_cast_fp16")];
+            tensor<fp16, [?, 400]> frames_9_cast_fp16 = sub(x = frames_7_cast_fp16, y = var_74_cast_fp16)[name = tensor<string, []>("frames_9_cast_fp16")];
+            tensor<fp16, [1, 400]> window_to_fp16 = const()[name = tensor<string, []>("window_to_fp16"), val = tensor<fp16, [1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320128)))];
+            tensor<fp16, [?, 400]> frames_11_cast_fp16 = mul(x = frames_9_cast_fp16, y = window_to_fp16)[name = tensor<string, []>("frames_11_cast_fp16")];
+            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [?, 1, 400]> input_cast_fp16 = expand_dims(axes = input_axes_0, x = frames_11_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [6]> var_85_pad_0 = const()[name = tensor<string, []>("op_85_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 112])];
+            tensor<string, []> var_85_mode_0 = const()[name = tensor<string, []>("op_85_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [?, 1, 512]> var_85_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = var_85_mode_0, pad = var_85_pad_0, x = input_cast_fp16)[name = tensor<string, []>("op_85_cast_fp16")];
+            tensor<string, []> var_105_pad_type_0 = const()[name = tensor<string, []>("op_105_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_105_strides_0 = const()[name = tensor<string, []>("op_105_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_105_pad_0 = const()[name = tensor<string, []>("op_105_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_105_dilations_0 = const()[name = tensor<string, []>("op_105_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_105_groups_0 = const()[name = tensor<string, []>("op_105_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> dft_real_weight_to_fp16 = const()[name = tensor<string, []>("dft_real_weight_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(321024)))];
+            tensor<fp16, [?, 257, 1]> var_105_cast_fp16 = conv(dilations = var_105_dilations_0, groups = var_105_groups_0, pad = var_105_pad_0, pad_type = var_105_pad_type_0, strides = var_105_strides_0, weight = dft_real_weight_to_fp16, x = var_85_cast_fp16)[name = tensor<string, []>("op_105_cast_fp16")];
+            tensor<int32, [1]> real_axes_0 = const()[name = tensor<string, []>("real_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [?, 257]> real_cast_fp16 = squeeze(axes = real_axes_0, x = var_105_cast_fp16)[name = tensor<string, []>("real_cast_fp16")];
+            tensor<string, []> var_123_pad_type_0 = const()[name = tensor<string, []>("op_123_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_123_strides_0 = const()[name = tensor<string, []>("op_123_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_123_pad_0 = const()[name = tensor<string, []>("op_123_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_123_dilations_0 = const()[name = tensor<string, []>("op_123_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_123_groups_0 = const()[name = tensor<string, []>("op_123_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> dft_imag_weight_to_fp16 = const()[name = tensor<string, []>("dft_imag_weight_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(584256)))];
+            tensor<fp16, [?, 257, 1]> var_123_cast_fp16 = conv(dilations = var_123_dilations_0, groups = var_123_groups_0, pad = var_123_pad_0, pad_type = var_123_pad_type_0, strides = var_123_strides_0, weight = dft_imag_weight_to_fp16, x = var_85_cast_fp16)[name = tensor<string, []>("op_123_cast_fp16")];
+            tensor<int32, [1]> imag_axes_0 = const()[name = tensor<string, []>("imag_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [?, 257]> imag_cast_fp16 = squeeze(axes = imag_axes_0, x = var_123_cast_fp16)[name = tensor<string, []>("imag_cast_fp16")];
+            tensor<fp16, []> var_126_promoted_to_fp16 = const()[name = tensor<string, []>("op_126_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [?, 257]> var_127_cast_fp16 = pow(x = real_cast_fp16, y = var_126_promoted_to_fp16)[name = tensor<string, []>("op_127_cast_fp16")];
+            tensor<fp16, []> var_128_promoted_to_fp16 = const()[name = tensor<string, []>("op_128_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [?, 257]> var_129_cast_fp16 = pow(x = imag_cast_fp16, y = var_128_promoted_to_fp16)[name = tensor<string, []>("op_129_cast_fp16")];
+            tensor<fp16, [?, 257]> power_cast_fp16 = add(x = var_127_cast_fp16, y = var_129_cast_fp16)[name = tensor<string, []>("power_cast_fp16")];
+            tensor<int32, [1]> var_133_axes_0 = const()[name = tensor<string, []>("op_133_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [?, 257, 1]> var_133_cast_fp16 = expand_dims(axes = var_133_axes_0, x = power_cast_fp16)[name = tensor<string, []>("op_133_cast_fp16")];
+            tensor<string, []> var_149_pad_type_0 = const()[name = tensor<string, []>("op_149_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_149_strides_0 = const()[name = tensor<string, []>("op_149_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_149_pad_0 = const()[name = tensor<string, []>("op_149_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_149_dilations_0 = const()[name = tensor<string, []>("op_149_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_149_groups_0 = const()[name = tensor<string, []>("op_149_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [80, 257, 1]> mel_weight_to_fp16 = const()[name = tensor<string, []>("mel_weight_to_fp16"), val = tensor<fp16, [80, 257, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(847488)))];
+            tensor<fp16, [?, 80, 1]> var_149_cast_fp16 = conv(dilations = var_149_dilations_0, groups = var_149_groups_0, pad = var_149_pad_0, pad_type = var_149_pad_type_0, strides = var_149_strides_0, weight = mel_weight_to_fp16, x = var_133_cast_fp16)[name = tensor<string, []>("op_149_cast_fp16")];
+            tensor<int32, [1]> mel_1_axes_0 = const()[name = tensor<string, []>("mel_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [?, 80]> mel_1_cast_fp16 = squeeze(axes = mel_1_axes_0, x = var_149_cast_fp16)[name = tensor<string, []>("mel_1_cast_fp16")];
+            tensor<fp16, []> eps_to_fp16 = const()[name = tensor<string, []>("eps_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [?, 80]> clip_0_cast_fp16 = clip(alpha = eps_to_fp16, beta = const_2_to_fp16, x = mel_1_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<string, []> clip_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("clip_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> mel_epsilon_0 = const()[name = tensor<string, []>("mel_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [?, 80]> clip_0_cast_fp16_to_fp32 = cast(dtype = clip_0_cast_fp16_to_fp32_dtype_0, x = clip_0_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [?, 80]> mel = log(epsilon = mel_epsilon_0, x = clip_0_cast_fp16_to_fp32)[name = tensor<string, []>("mel")];
+            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([-1, 998, 80])];
+            tensor<string, []> mel_to_fp16_dtype_0 = const()[name = tensor<string, []>("mel_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [?, 80]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [?, 998, 80]> var_157_cast_fp16 = reshape(shape = concat_1x, x = mel_to_fp16)[name = tensor<string, []>("op_157_cast_fp16")];
+            tensor<string, []> var_157_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_157_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [?, 998, 80]> fbank = cast(dtype = var_157_cast_fp16_to_fp32_dtype_0, x = var_157_cast_fp16)[name = tensor<string, []>("cast_2")];
+        } -> (fbank);
+}

FBank.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d992fbcd8d26540cfcb291d86417bf9bd2c94ac15295c8ff70b3b93ccd5158ed
+size 888672