Update Melspectrogram interface

nvidia_parakeet-v2/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f9e0ad4bff30b21d884c2cdc5bf1551b59209d9314c93a11f6bc4a21e4d26b5
+oid sha256:9e6faff9fec59f231371b1fffcdd7a9f50231cff77aa1744af91907437bc30f6
 size 243

nvidia_parakeet-v2/MelSpectrogram.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1ff0c946a5a04b1f3d754001c0951bea58f79e57e04124d6da0f03a09c535acd
-size 327
+oid sha256:623b0eab613190f6748463347d4cf1980003c067b0ad0d9cdcdac0de231e3ff0
+size 329

nvidia_parakeet-v2/MelSpectrogram.mlmodelc/metadata.json

@@ -7,9 +7,9 @@
         "hasShapeFlexibility" : "0",
         "isOptional" : "0",
         "dataType" : "Float16",
-        "formattedType" : "MultiArray (Float16 128 × 1501)",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1501 × 128)",
         "shortDescription" : "",
-        "shape" : "[128, 1501]",
+        "shape" : "[1, 1, 1501, 128]",
         "name" : "melspectrogram_features",
         "type" : "MultiArray"
       }
@@ -22,6 +22,7 @@
       "Ios17.mul" : 2,
       "Ios17.sqrt" : 1,
       "Ios17.square" : 3,
+      "Ios17.transpose" : 1,
       "Ios17.sub" : 2,
       "Ios17.matmul" : 1,
       "Ios17.conv" : 2,
@@ -30,7 +31,7 @@
       "Ios17.add" : 3,
       "Ios16.reduceMean" : 2,
       "Ios17.realDiv" : 1,
-      "Ios17.expandDims" : 2,
+      "Ios17.expandDims" : 4,
       "Ios17.squeeze" : 2,
       "Ios17.reshape" : 2,
       "Identity" : 1,
@@ -54,8 +55,8 @@
     },
     "userDefinedMetadata" : {
       "com.github.apple.coremltools.source_dialect" : "TorchScript",
-      "com.github.apple.coremltools.version" : "8.2",
-      "com.github.apple.coremltools.source" : "torch==2.6.0"
+      "com.github.apple.coremltools.source" : "torch==2.6.0",
+      "com.github.apple.coremltools.version" : "8.2"
     },
     "inputSchema" : [
       {

nvidia_parakeet-v2/MelSpectrogram.mlmodelc/model.mil

@@ -56,12 +56,12 @@ program(1.0)
             tensor<fp16, [128, 1501]> mel_spec_1_cast_fp16 = matmul(transpose_x = mel_spec_1_transpose_x_0, transpose_y = mel_spec_1_transpose_y_0, x = mel_filters_to_fp16, y = magnitudes_cast_fp16)[name = tensor<string, []>("mel_spec_1_cast_fp16")];
             tensor<fp16, []> var_56_to_fp16 = const()[name = tensor<string, []>("op_56_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
             tensor<fp16, [128, 1501]> mel_spec_3_cast_fp16 = add(x = mel_spec_1_cast_fp16, y = var_56_to_fp16)[name = tensor<string, []>("mel_spec_3_cast_fp16")];
-            tensor<fp32, []> mel_spec_epsilon_0 = const()[name = tensor<string, []>("mel_spec_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
-            tensor<fp16, [128, 1501]> mel_spec_cast_fp16 = log(epsilon = mel_spec_epsilon_0, x = mel_spec_3_cast_fp16)[name = tensor<string, []>("mel_spec_cast_fp16")];
+            tensor<fp32, []> mel_spec_5_epsilon_0 = const()[name = tensor<string, []>("mel_spec_5_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [128, 1501]> mel_spec_5_cast_fp16 = log(epsilon = mel_spec_5_epsilon_0, x = mel_spec_3_cast_fp16)[name = tensor<string, []>("mel_spec_5_cast_fp16")];
             tensor<int32, [1]> per_feature_mean_axes_0 = const()[name = tensor<string, []>("per_feature_mean_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<bool, []> per_feature_mean_keep_dims_0 = const()[name = tensor<string, []>("per_feature_mean_keep_dims_0"), val = tensor<bool, []>(true)];
-            tensor<fp16, [128, 1]> per_feature_mean_cast_fp16 = reduce_mean(axes = per_feature_mean_axes_0, keep_dims = per_feature_mean_keep_dims_0, x = mel_spec_cast_fp16)[name = tensor<string, []>("per_feature_mean_cast_fp16")];
-            tensor<fp16, [128, 1501]> sub_0_cast_fp16 = sub(x = mel_spec_cast_fp16, y = per_feature_mean_cast_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
+            tensor<fp16, [128, 1]> per_feature_mean_cast_fp16 = reduce_mean(axes = per_feature_mean_axes_0, keep_dims = per_feature_mean_keep_dims_0, x = mel_spec_5_cast_fp16)[name = tensor<string, []>("per_feature_mean_cast_fp16")];
+            tensor<fp16, [128, 1501]> sub_0_cast_fp16 = sub(x = mel_spec_5_cast_fp16, y = per_feature_mean_cast_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
             tensor<fp16, [128, 1501]> square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor<string, []>("square_0_cast_fp16")];
             tensor<int32, [1]> reduce_mean_1_axes_0 = const()[name = tensor<string, []>("reduce_mean_1_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<bool, []> reduce_mean_1_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_1_keep_dims_0"), val = tensor<bool, []>(true)];
@@ -71,6 +71,12 @@ program(1.0)
             tensor<fp16, [128, 1]> sqrt_0_cast_fp16 = sqrt(x = mul_0_cast_fp16)[name = tensor<string, []>("sqrt_0_cast_fp16")];
             tensor<fp16, []> var_70_to_fp16 = const()[name = tensor<string, []>("op_70_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
             tensor<fp16, [128, 1]> per_feature_std_cast_fp16 = add(x = sqrt_0_cast_fp16, y = var_70_to_fp16)[name = tensor<string, []>("per_feature_std_cast_fp16")];
-            tensor<fp16, [128, 1501]> melspectrogram_features = real_div(x = sub_0_cast_fp16, y = per_feature_std_cast_fp16)[name = tensor<string, []>("op_74_cast_fp16")];
+            tensor<fp16, [128, 1501]> mel_spec_cast_fp16 = real_div(x = sub_0_cast_fp16, y = per_feature_std_cast_fp16)[name = tensor<string, []>("mel_spec_cast_fp16")];
+            tensor<int32, [2]> var_75_perm_0 = const()[name = tensor<string, []>("op_75_perm_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<int32, [1]> var_77_axes_0 = const()[name = tensor<string, []>("op_77_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1501, 128]> var_75_cast_fp16 = transpose(perm = var_75_perm_0, x = mel_spec_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1501, 128]> var_77_cast_fp16 = expand_dims(axes = var_77_axes_0, x = var_75_cast_fp16)[name = tensor<string, []>("op_77_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = tensor<string, []>("op_79_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1501, 128]> melspectrogram_features = expand_dims(axes = var_79_axes_0, x = var_77_cast_fp16)[name = tensor<string, []>("op_79_cast_fp16")];
         } -> (melspectrogram_features);
 }