Upload Parakeet-TDT v3 CoreML models

Decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9a7c21bb27b230d36103438ceff23bec13c018989b66c3bea507a4c6869385b
+size 243

Decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2652a93062cfc1db8675cf8e9246ff72bf92754f44ea402815373232a92488c
+size 554

Decoder.mlmodelc/metadata.json

@@ -0,0 +1,122 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 4,
+      "Ios17.gather" : 1,
+      "Ios17.lstm" : 2,
+      "Identity" : 1,
+      "Ios17.transpose" : 2,
+      "Split" : 2,
+      "Ios17.add" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Stack" : 2
+    },
+    "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" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-02-12",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_decoder",
+    "method" : "predict"
+  }
+]

Decoder.mlmodelc/model.mil

@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<fp32, [8193, 640]> module_prediction_embed_weight = const()[name = tensor<string, []>("module_prediction_embed_weight"), val = tensor<fp32, [8193, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = targets, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(8193)];
+            tensor<int32, [1, 1]> add_2 = add(x = targets, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = targets, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_axis_1 = const()[name = tensor<string, []>("y_axis_1"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> y = gather(axis = y_axis_1, batch_dims = y_batch_dims_0, indices = select_0, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight)[name = tensor<string, []>("y")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_0_0, tensor<fp32, [1, 1, 640]> split_0_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in)[name = tensor<string, []>("split_0")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_1_0, tensor<fp32, [1, 1, 640]> split_1_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in)[name = tensor<string, []>("split_1")];
+            tensor<fp32, [2560]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20974208)))];
+            tensor<fp32, [2560, 640]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20984512)))];
+            tensor<fp32, [2560, 640]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27538176)))];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = y)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 1, 640]> input_lstm_layer_0_0, tensor<fp32, [1, 640]> input_lstm_layer_0_1, tensor<fp32, [1, 640]> input_lstm_layer_0_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze, initial_h = input_lstm_layer_0_lstm_h0_squeeze, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2, weight_ih = concat_1, x = input_3)[name = tensor<string, []>("input_lstm_layer_0")];
+            tensor<fp32, [2560]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34091840)))];
+            tensor<fp32, [2560, 640]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34102144)))];
+            tensor<fp32, [2560, 640]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40655808)))];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_h0_squeeze = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_1)[name = tensor<string, []>("input_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_c0_squeeze = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_1)[name = tensor<string, []>("input_lstm_c0_squeeze")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_0, tensor<fp32, [1, 640]> input_1, tensor<fp32, [1, 640]> input_2 = lstm(activation = input_activation_0, bias = concat_3, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze, initial_h = input_lstm_h0_squeeze, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5, weight_ih = concat_4, x = input_lstm_layer_0_0)[name = tensor<string, []>("input")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> h_out = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_1, input_1))[name = tensor<string, []>("obj_3")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> c_out = stack(axis = obj_axis_0, values = (input_lstm_layer_0_2, input_2))[name = tensor<string, []>("obj")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<fp32, [1, 640, 1]> decoder = transpose(perm = transpose_0_perm_0, x = input_0)[name = tensor<string, []>("transpose_1")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}

Decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90845baea25befede87e504148494a7de982e889ae979414ec62df1cebad67c8
+size 47209472

Encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77cc560a12972e8aa4d389042e30d373fc3df38ab1b4a7016ded4fc2e63f3e19
+size 243

Encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:986ae7bd544c51f342203afe6b7b086053256479d193601c937221b0ef7a33c1
+size 478

Encoder.mlmodelc/metadata.json

@@ -0,0 +1,105 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet encoder (15 s window)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 188)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 188]",
+        "name" : "encoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoder_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.floor" : 3,
+      "Ios17.logicalAnd" : 2,
+      "Ios17.reshape" : 145,
+      "Ios16.softmax" : 24,
+      "Ios17.matmul" : 72,
+      "Ios17.transpose" : 172,
+      "Split" : 24,
+      "Select" : 72,
+      "Ios17.expandDims" : 5,
+      "Ios17.add" : 174,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 48,
+      "Ios16.sigmoid" : 24,
+      "Pad" : 48,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 120,
+      "Ios16.silu" : 72,
+      "Ios17.less" : 1,
+      "Ios17.conv" : 77,
+      "Ios16.relu" : 3,
+      "Ios17.cast" : 4,
+      "Ios17.linear" : 193,
+      "Ios17.mul" : 99
+    },
+    "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 × 128 × 1501)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1501]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-02-12",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_encoder",
+    "method" : "predict"
+  }
+]

Encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9dbf2f3e759e7ad1d88dda24cc4545ce96bbbe2f778d8b3f4d4393fa51afabd
+size 1185860032

JointDecision.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8d149ecb731de58bd4bb7ebc97f0b5ce8da6186b23bccbb556411b08ddc24ca
+size 243

JointDecision.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18a2c6d0b46516234c4e929cc44178255b47e79d5e28dc3f96726a249272f25e
+size 591

JointDecision.mlmodelc/metadata.json

@@ -0,0 +1,123 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet single-step joint decision (current frame)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_id",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_prob",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "duration",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reduceArgmax" : 2,
+      "Ios17.squeeze" : 1,
+      "Ios17.linear" : 3,
+      "Ios16.relu" : 1,
+      "Ios17.transpose" : 2,
+      "Ios17.sliceByIndex" : 2,
+      "Ios17.add" : 1,
+      "Ios16.softmax" : 1,
+      "Ios17.gatherAlongAxis" : 1,
+      "Ios17.topk" : 1,
+      "Ios17.expandDims" : 3
+    },
+    "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" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 1]",
+        "name" : "encoder_step",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_step",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-02-12",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_joint_decision_single_step",
+    "method" : "predict"
+  }
+]

JointDecision.mlmodelc/model.mil

@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder_step, tensor<fp32, [1, 1024, 1]> encoder_step) {
+            tensor<fp32, [640]> joint_module_enc_bias = const()[name = tensor<string, []>("joint_module_enc_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [640, 1024]> joint_module_enc_weight = const()[name = tensor<string, []>("joint_module_enc_weight"), val = tensor<fp32, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2688)))];
+            tensor<fp32, [640]> joint_module_pred_bias = const()[name = tensor<string, []>("joint_module_pred_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624192)))];
+            tensor<fp32, [640, 640]> joint_module_pred_weight = const()[name = tensor<string, []>("joint_module_pred_weight"), val = tensor<fp32, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2626816)))];
+            tensor<fp32, [8198]> joint_module_joint_net_2_bias = const()[name = tensor<string, []>("joint_module_joint_net_2_bias"), val = tensor<fp32, [8198]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4265280)))];
+            tensor<fp32, [8198, 640]> joint_module_joint_net_2_weight = const()[name = tensor<string, []>("joint_module_joint_net_2_weight"), val = tensor<fp32, [8198, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4298176)))];
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1, 1024]> input_1 = transpose(perm = input_1_perm_0, x = encoder_step)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 1, 640]> enc_proj = linear(bias = joint_module_enc_bias, weight = joint_module_enc_weight, x = input_1)[name = tensor<string, []>("linear_0")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = decoder_step)[name = tensor<string, []>("transpose_0")];
+            tensor<fp32, [1, 1, 640]> dec_proj = linear(bias = joint_module_pred_bias, weight = joint_module_pred_weight, x = input_3)[name = tensor<string, []>("linear_1")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 1, 1, 640]> var_23 = expand_dims(axes = var_23_axes_0, x = enc_proj)[name = tensor<string, []>("op_23")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 1, 640]> var_24 = expand_dims(axes = var_24_axes_0, x = dec_proj)[name = tensor<string, []>("op_24")];
+            tensor<fp32, [1, 1, 1, 640]> input_5 = add(x = var_23, y = var_24)[name = tensor<string, []>("input_5")];
+            tensor<fp32, [1, 1, 1, 640]> input_7 = relu(x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 1, 1, 8198]> logits = linear(bias = joint_module_joint_net_2_bias, weight = joint_module_joint_net_2_weight, x = input_7)[name = tensor<string, []>("linear_2")];
+            tensor<int32, [4]> token_logits_begin_0 = const()[name = tensor<string, []>("token_logits_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> token_logits_end_0 = const()[name = tensor<string, []>("token_logits_end_0"), val = tensor<int32, [4]>([1, 1, 1, 8193])];
+            tensor<bool, [4]> token_logits_end_mask_0 = const()[name = tensor<string, []>("token_logits_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 1, 1, 8193]> token_logits = slice_by_index(begin = token_logits_begin_0, end = token_logits_end_0, end_mask = token_logits_end_mask_0, x = logits)[name = tensor<string, []>("token_logits")];
+            tensor<int32, [4]> duration_logits_begin_0 = const()[name = tensor<string, []>("duration_logits_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 8193])];
+            tensor<int32, [4]> duration_logits_end_0 = const()[name = tensor<string, []>("duration_logits_end_0"), val = tensor<int32, [4]>([1, 1, 1, 8198])];
+            tensor<bool, [4]> duration_logits_end_mask_0 = const()[name = tensor<string, []>("duration_logits_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 1, 1, 5]> duration_logits = slice_by_index(begin = duration_logits_begin_0, end = duration_logits_end_0, end_mask = duration_logits_end_mask_0, x = logits)[name = tensor<string, []>("duration_logits")];
+            tensor<int32, []> var_43_axis_0 = const()[name = tensor<string, []>("op_43_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_43_keep_dims_0 = const()[name = tensor<string, []>("op_43_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_43_output_dtype_0 = const()[name = tensor<string, []>("op_43_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> token_id = reduce_argmax(axis = var_43_axis_0, keep_dims = var_43_keep_dims_0, output_dtype = var_43_output_dtype_0, x = token_logits)[name = tensor<string, []>("op_43")];
+            tensor<int32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<int32, []>(-1)];
+            tensor<fp32, [1, 1, 1, 8193]> token_probs_all = softmax(axis = var_49, x = token_logits)[name = tensor<string, []>("token_probs_all")];
+            tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1, 1, 1]> var_58 = expand_dims(axes = var_58_axes_0, x = token_id)[name = tensor<string, []>("op_58")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_61_validate_indices_0 = const()[name = tensor<string, []>("op_61_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1, 1, 1]> var_61 = gather_along_axis(axis = var_59, indices = var_58, validate_indices = var_61_validate_indices_0, x = token_probs_all)[name = tensor<string, []>("op_61")];
+            tensor<int32, [1]> var_63_axes_0 = const()[name = tensor<string, []>("op_63_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 1, 1]> token_prob = squeeze(axes = var_63_axes_0, x = var_61)[name = tensor<string, []>("op_63")];
+            tensor<int32, []> var_66_axis_0 = const()[name = tensor<string, []>("op_66_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_66_keep_dims_0 = const()[name = tensor<string, []>("op_66_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_66_output_dtype_0 = const()[name = tensor<string, []>("op_66_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> duration = reduce_argmax(axis = var_66_axis_0, keep_dims = var_66_keep_dims_0, output_dtype = var_66_output_dtype_0, x = duration_logits)[name = tensor<string, []>("op_66")];
+            tensor<int32, []> var_72 = const()[name = tensor<string, []>("op_72"), val = tensor<int32, []>(64)];
+            tensor<int32, []> var_76_axis_0 = const()[name = tensor<string, []>("op_76_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_76_ascending_0 = const()[name = tensor<string, []>("op_76_ascending_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_76_sort_0 = const()[name = tensor<string, []>("op_76_sort_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_76_return_indices_0 = const()[name = tensor<string, []>("op_76_return_indices_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> var_76_output_indices_dtype_0 = const()[name = tensor<string, []>("op_76_output_indices_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits, tensor<int32, [1, 1, 1, 64]> top_k_ids = topk(ascending = var_76_ascending_0, axis = var_76_axis_0, k = var_72, output_indices_dtype = var_76_output_indices_dtype_0, return_indices = var_76_return_indices_0, sort = var_76_sort_0, x = token_logits)[name = tensor<string, []>("op_76")];
+        } -> (token_id, token_prob, duration, top_k_ids, top_k_logits);
+}

JointDecision.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fe093ad8a2e61ba992a6c94ce5f281203340f3ff64ca86bee8cb468d7f4bba8
+size 25285120

Preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ec76b3a51582dce0f996aefeb0d4ae42436851a851372f7c42c74007c418306
+size 243

Preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd4c07458015c6c12e5e9526e3f002789302d78569601bd985a2f882f34bb473
+size 493

Preprocessor.mlmodelc/metadata.json

@@ -0,0 +1,110 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet preprocessor (15 s window)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float32",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 2,
+      "Ios17.reshape" : 2,
+      "Identity" : 1,
+      "Ios17.matmul" : 1,
+      "Ios17.expandDims" : 10,
+      "Select" : 3,
+      "Ios17.add" : 4,
+      "Tile" : 2,
+      "Ios17.sliceByIndex" : 3,
+      "Ios16.reduceSum" : 4,
+      "Shape" : 3,
+      "Ios17.gather" : 3,
+      "Pad" : 1,
+      "Ios17.log" : 1,
+      "Ios17.conv" : 2,
+      "Ios17.sub" : 4,
+      "Ios17.pow" : 2,
+      "Ios17.cast" : 3,
+      "Ios17.realDiv" : 4,
+      "Stack" : 1,
+      "Ios17.concat" : 3,
+      "Ios17.floorDiv" : 1,
+      "Ios17.less" : 1,
+      "Ios17.sqrt" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.mul" : 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 × 1...240000",
+        "shapeRange" : "[[1, 1], [1, 240000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "audio_signal",
+        "shortDescription" : ""
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-02-12",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_preprocessor",
+    "method" : "predict"
+  }
+]

Preprocessor.mlmodelc/model.mil

@@ -0,0 +1,156 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<int32, [1]> audio_length, tensor<fp32, [1, ?]> audio_signal) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio_signal", [1, 1]}}), ("RangeDims", {{"audio_signal", [[1, 1], [1, 240000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<fp32, []> var_24 = const()[name = tensor<string, []>("op_24"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, []> var_25 = const()[name = tensor<string, []>("op_25"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
+            tensor<int32, []> var_34 = const()[name = tensor<string, []>("op_34"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_35 = add(x = audio_length, y = var_34)[name = tensor<string, []>("op_35")];
+            tensor<int32, []> var_36 = const()[name = tensor<string, []>("op_36"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_37 = sub(x = var_35, y = var_36)[name = tensor<string, []>("op_37")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_37, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<string, []> var_38_dtype_0 = const()[name = tensor<string, []>("op_38_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> var_39_promoted = const()[name = tensor<string, []>("op_39_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_38 = cast(dtype = var_38_dtype_0, x = floor_div_0)[name = tensor<string, []>("cast_13")];
+            tensor<fp32, [1]> seq_len_1 = add(x = var_38, y = var_39_promoted)[name = tensor<string, []>("seq_len_1")];
+            tensor<string, []> seq_len_dtype_0 = const()[name = tensor<string, []>("seq_len_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [2]> var_43_begin_0 = const()[name = tensor<string, []>("op_43_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_43_end_0 = const()[name = tensor<string, []>("op_43_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_43_end_mask_0 = const()[name = tensor<string, []>("op_43_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_43_squeeze_mask_0 = const()[name = tensor<string, []>("op_43_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp32, [1]> var_43 = slice_by_index(begin = var_43_begin_0, end = var_43_end_0, end_mask = var_43_end_mask_0, squeeze_mask = var_43_squeeze_mask_0, x = audio_signal)[name = tensor<string, []>("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = tensor<string, []>("op_44")];
+            tensor<int32, [2]> var_46_begin_0 = const()[name = tensor<string, []>("op_46_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_46_end_0 = const()[name = tensor<string, []>("op_46_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_46_end_mask_0 = const()[name = tensor<string, []>("op_46_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp32, [1, ?]> var_46 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_46")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = tensor<string, []>("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = tensor<string, []>("op_48_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = tensor<string, []>("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [1, ?]> var_48 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, x = audio_signal)[name = tensor<string, []>("op_48")];
+            tensor<fp32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<fp32, []>(0x1.f0a3d8p-1)];
+            tensor<fp32, [1, ?]> var_50 = mul(x = var_48, y = var_49)[name = tensor<string, []>("op_50")];
+            tensor<fp32, [1, ?]> var_51 = sub(x = var_46, y = var_50)[name = tensor<string, []>("op_51")];
+            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, ?]> input_1 = concat(axis = var_9, interleave = input_1_interleave_0, values = (var_44, var_51))[name = tensor<string, []>("input_1")];
+            tensor<int32, [3]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp32, [1, 1, ?]> input_3 = reshape(shape = concat_0x, x = input_1)[name = tensor<string, []>("input_3")];
+            tensor<fp32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<fp32, [1, 1, ?]> input_5 = pad(constant_val = const_1, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3)[name = tensor<string, []>("input_5")];
+            tensor<int32, [2]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp32, [1, ?]> input = reshape(shape = concat_1x, x = input_5)[name = tensor<string, []>("input")];
+            tensor<fp32, [257, 1, 512]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [257, 1, 512]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526464)))];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> expand_dims_4 = expand_dims(axes = expand_dims_4_axes_0, x = input)[name = tensor<string, []>("expand_dims_4")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_0 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1, x = expand_dims_4)[name = tensor<string, []>("conv_0")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_1 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2, x = expand_dims_4)[name = tensor<string, []>("conv_1")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp32, [1, 257, ?, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (conv_0, conv_1))[name = tensor<string, []>("stack_0")];
+            tensor<fp32, []> var_17_promoted = const()[name = tensor<string, []>("op_17_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, 257, ?, 2]> var_67 = pow(x = stack_0, y = var_17_promoted)[name = tensor<string, []>("op_67")];
+            tensor<int32, [1]> var_69_axes_0 = const()[name = tensor<string, []>("op_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_69_keep_dims_0 = const()[name = tensor<string, []>("op_69_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 257, ?]> var_69 = reduce_sum(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_67)[name = tensor<string, []>("op_69")];
+            tensor<fp32, [1, 257, ?]> x_9 = identity(x = var_69)[name = tensor<string, []>("x_9")];
+            tensor<fp32, [1, 128, 257]> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<fp32, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1052864)))];
+            tensor<bool, []> x_11_transpose_x_0 = const()[name = tensor<string, []>("x_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_11_transpose_y_0 = const()[name = tensor<string, []>("x_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 128, ?]> x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = const_2, y = x_9)[name = tensor<string, []>("x_11")];
+            tensor<fp32, []> var_76 = const()[name = tensor<string, []>("op_76"), val = tensor<fp32, []>(0x1p-24)];
+            tensor<fp32, [1, 128, ?]> var_77 = add(x = x_11, y = var_76)[name = tensor<string, []>("op_77")];
+            tensor<fp32, []> x_13_epsilon_0 = const()[name = tensor<string, []>("x_13_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [1, 128, ?]> x_13 = log(epsilon = x_13_epsilon_0, x = var_77)[name = tensor<string, []>("x_13")];
+            tensor<int32, [3]> var_79_shape = shape(x = x_13)[name = tensor<string, []>("op_79_shape")];
+            tensor<int32, []> gather_4 = const()[name = tensor<string, []>("gather_4"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_5_batch_dims_0 = const()[name = tensor<string, []>("gather_5_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_5_validate_indices_0 = const()[name = tensor<string, []>("gather_5_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_2 = const()[name = tensor<string, []>("select_2"), val = tensor<int32, []>(2)];
+            tensor<int32, []> gather_5_axis_1 = const()[name = tensor<string, []>("gather_5_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_5 = gather(axis = gather_5_axis_1, batch_dims = gather_5_batch_dims_0, indices = select_2, validate_indices = gather_5_validate_indices_0, x = var_79_shape)[name = tensor<string, []>("gather_5")];
+            tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> var_81 = range_1d(end = gather_5, start = const_3, step = const_4)[name = tensor<string, []>("op_81")];
+            tensor<int32, [1]> var_82_axes_0 = const()[name = tensor<string, []>("op_82_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_82 = expand_dims(axes = var_82_axes_0, x = var_81)[name = tensor<string, []>("op_82")];
+            tensor<int32, []> concat_2_axis_0 = const()[name = tensor<string, []>("concat_2_axis_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> concat_2_interleave_0 = const()[name = tensor<string, []>("concat_2_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [2]> concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (gather_4, gather_5))[name = tensor<string, []>("concat_2")];
+            tensor<int32, [2]> shape_0 = shape(x = var_82)[name = tensor<string, []>("shape_0")];
+            tensor<int32, [2]> real_div_0 = real_div(x = concat_2, y = shape_0)[name = tensor<string, []>("real_div_0")];
+            tensor<int32, [?, ?]> time_steps = tile(reps = real_div_0, x = var_82)[name = tensor<string, []>("time_steps")];
+            tensor<int32, [1]> var_85_axes_0 = const()[name = tensor<string, []>("op_85_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> mel_length = cast(dtype = seq_len_dtype_0, x = seq_len_1)[name = tensor<string, []>("cast_12")];
+            tensor<int32, [1, 1]> var_85 = expand_dims(axes = var_85_axes_0, x = mel_length)[name = tensor<string, []>("op_85")];
+            tensor<bool, [?, ?]> valid_mask = less(x = time_steps, y = var_85)[name = tensor<string, []>("valid_mask")];
+            tensor<int32, [1]> var_87_axes_0 = const()[name = tensor<string, []>("op_87_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [?, 1, ?]> var_87 = expand_dims(axes = var_87_axes_0, x = valid_mask)[name = tensor<string, []>("op_87")];
+            tensor<fp32, [1, 128, ?]> var_88 = select(a = x_13, b = var_24, cond = var_87)[name = tensor<string, []>("op_88")];
+            tensor<int32, [1]> x_mean_numerator_axes_0 = const()[name = tensor<string, []>("x_mean_numerator_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<bool, []> x_mean_numerator_keep_dims_0 = const()[name = tensor<string, []>("x_mean_numerator_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 128]> x_mean_numerator = reduce_sum(axes = x_mean_numerator_axes_0, keep_dims = x_mean_numerator_keep_dims_0, x = var_88)[name = tensor<string, []>("x_mean_numerator")];
+            tensor<string, []> cast_3_dtype_0 = const()[name = tensor<string, []>("cast_3_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> x_mean_denominator_axes_0 = const()[name = tensor<string, []>("x_mean_denominator_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> x_mean_denominator_keep_dims_0 = const()[name = tensor<string, []>("x_mean_denominator_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [?, ?]> cast_3 = cast(dtype = cast_3_dtype_0, x = valid_mask)[name = tensor<string, []>("cast_11")];
+            tensor<fp32, [?]> x_mean_denominator = reduce_sum(axes = x_mean_denominator_axes_0, keep_dims = x_mean_denominator_keep_dims_0, x = cast_3)[name = tensor<string, []>("x_mean_denominator")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [?, 1]> var_93 = expand_dims(axes = var_93_axes_0, x = x_mean_denominator)[name = tensor<string, []>("op_93")];
+            tensor<fp32, [?, 128]> x_mean = real_div(x = x_mean_numerator, y = var_93)[name = tensor<string, []>("x_mean")];
+            tensor<int32, [1]> var_96_axes_0 = const()[name = tensor<string, []>("op_96_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [?, 128, 1]> var_96 = expand_dims(axes = var_96_axes_0, x = x_mean)[name = tensor<string, []>("op_96")];
+            tensor<fp32, [?, 128, ?]> var_97 = sub(x = x_13, y = var_96)[name = tensor<string, []>("op_97")];
+            tensor<fp32, [?, 128, ?]> var_98 = select(a = var_97, b = var_24, cond = var_87)[name = tensor<string, []>("op_98")];
+            tensor<fp32, []> var_17_promoted_1 = const()[name = tensor<string, []>("op_17_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [?, 128, ?]> var_99 = pow(x = var_98, y = var_17_promoted_1)[name = tensor<string, []>("op_99")];
+            tensor<int32, [1]> var_101_axes_0 = const()[name = tensor<string, []>("op_101_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<bool, []> var_101_keep_dims_0 = const()[name = tensor<string, []>("op_101_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [?, 128]> var_101 = reduce_sum(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_99)[name = tensor<string, []>("op_101")];
+            tensor<fp32, []> var_103 = const()[name = tensor<string, []>("op_103"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [?, 1]> var_104 = sub(x = var_93, y = var_103)[name = tensor<string, []>("op_104")];
+            tensor<fp32, [?, 128]> var_105 = real_div(x = var_101, y = var_104)[name = tensor<string, []>("op_105")];
+            tensor<fp32, [?, 128]> x_std_1 = sqrt(x = var_105)[name = tensor<string, []>("x_std_1")];
+            tensor<fp32, [?, 128]> x_std = add(x = x_std_1, y = var_25)[name = tensor<string, []>("x_std")];
+            tensor<int32, [1]> var_110_axes_0 = const()[name = tensor<string, []>("op_110_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [?, 128, 1]> var_110 = expand_dims(axes = var_110_axes_0, x = x_std)[name = tensor<string, []>("op_110")];
+            tensor<fp32, [?, 128, ?]> x = real_div(x = var_97, y = var_110)[name = tensor<string, []>("x")];
+            tensor<int32, [3]> var_112_shape = shape(x = x)[name = tensor<string, []>("op_112_shape")];
+            tensor<int32, []> gather_6_batch_dims_0 = const()[name = tensor<string, []>("gather_6_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_6_validate_indices_0 = const()[name = tensor<string, []>("gather_6_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_3 = const()[name = tensor<string, []>("select_3"), val = tensor<int32, []>(2)];
+            tensor<int32, []> gather_6_axis_1 = const()[name = tensor<string, []>("gather_6_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_6 = gather(axis = gather_6_axis_1, batch_dims = gather_6_batch_dims_0, indices = select_3, validate_indices = gather_6_validate_indices_0, x = var_112_shape)[name = tensor<string, []>("gather_6")];
+            tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_6, start = const_5, step = const_6)[name = tensor<string, []>("mask_1")];
+            tensor<int32, []> gather_7_batch_dims_0 = const()[name = tensor<string, []>("gather_7_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_7_validate_indices_0 = const()[name = tensor<string, []>("gather_7_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_4 = const()[name = tensor<string, []>("select_4"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_7_axis_1 = const()[name = tensor<string, []>("gather_7_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_7 = gather(axis = gather_7_axis_1, batch_dims = gather_7_batch_dims_0, indices = select_4, validate_indices = gather_7_validate_indices_0, x = var_112_shape)[name = tensor<string, []>("gather_7")];
+            tensor<int32, []> concat_3_axis_0 = const()[name = tensor<string, []>("concat_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> concat_3_interleave_0 = const()[name = tensor<string, []>("concat_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [2]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (gather_7, var_9))[name = tensor<string, []>("concat_3")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [?, ?]> var_116 = tile(reps = concat_3, x = expand_dims_0)[name = tensor<string, []>("op_116")];
+            tensor<bool, [?, ?]> mask = greater_equal(x = var_116, y = var_85)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_119_axes_0 = const()[name = tensor<string, []>("op_119_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [?, 1, ?]> var_119 = expand_dims(axes = var_119_axes_0, x = mask)[name = tensor<string, []>("op_119")];
+            tensor<fp32, [?, 128, ?]> mel = select(a = var_24, b = x, cond = var_119)[name = tensor<string, []>("processed_signal")];
+        } -> (mel, mel_length);
+}

Preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a31e2bd311bfa97ea3fe73d4a6c029824327a71590397bd8784089a88754a77
+size 1184512

metadata.json

@@ -0,0 +1,212 @@
+{
+  "model_id": "nvidia/parakeet-tdt-0.6b-v3",
+  "sample_rate": 16000,
+  "max_audio_seconds": 15.0,
+  "max_audio_samples": 240000,
+  "max_symbol_steps": 1,
+  "vocab_size": 8192,
+  "joint_extra_outputs": 5,
+  "checkpoint": {
+    "type": "pretrained",
+    "model_id": "nvidia/parakeet-tdt-0.6b-v3"
+  },
+  "coreml": {
+    "compute_units": "CPU_ONLY",
+    "compute_precision": "FLOAT32"
+  },
+  "components": {
+    "preprocessor": {
+      "inputs": {
+        "audio_signal": [
+          1,
+          240000
+        ],
+        "audio_length": [
+          1
+        ]
+      },
+      "outputs": {
+        "mel": [
+          1,
+          128,
+          1501
+        ],
+        "mel_length": [
+          1
+        ]
+      },
+      "path": "parakeet_preprocessor.mlpackage"
+    },
+    "encoder": {
+      "inputs": {
+        "mel": [
+          1,
+          128,
+          1501
+        ],
+        "mel_length": [
+          1
+        ]
+      },
+      "outputs": {
+        "encoder": [
+          1,
+          1024,
+          188
+        ],
+        "encoder_length": [
+          1
+        ]
+      },
+      "path": "parakeet_encoder.mlpackage"
+    },
+    "mel_encoder": {
+      "inputs": {
+        "audio_signal": [
+          1,
+          240000
+        ],
+        "audio_length": [
+          1
+        ]
+      },
+      "outputs": {
+        "encoder": [
+          1,
+          1024,
+          188
+        ],
+        "encoder_length": [
+          1
+        ]
+      },
+      "path": "parakeet_mel_encoder.mlpackage"
+    },
+    "decoder": {
+      "inputs": {
+        "targets": [
+          1,
+          1
+        ],
+        "target_length": [
+          1
+        ],
+        "h_in": [
+          2,
+          1,
+          640
+        ],
+        "c_in": [
+          2,
+          1,
+          640
+        ]
+      },
+      "outputs": {
+        "decoder": [
+          1,
+          640,
+          1
+        ],
+        "h_out": [
+          2,
+          1,
+          640
+        ],
+        "c_out": [
+          2,
+          1,
+          640
+        ]
+      },
+      "path": "parakeet_decoder.mlpackage"
+    },
+    "joint": {
+      "inputs": {
+        "encoder": [
+          1,
+          1024,
+          188
+        ],
+        "decoder": [
+          1,
+          640,
+          1
+        ]
+      },
+      "outputs": {
+        "logits": [
+          1,
+          188,
+          1,
+          8198
+        ]
+      },
+      "path": "parakeet_joint.mlpackage"
+    },
+    "joint_decision": {
+      "inputs": {
+        "encoder": [
+          1,
+          1024,
+          188
+        ],
+        "decoder": [
+          1,
+          640,
+          1
+        ]
+      },
+      "outputs": {
+        "token_id": [
+          1,
+          1024,
+          640
+        ],
+        "token_prob": [
+          1,
+          1024,
+          640
+        ],
+        "duration": [
+          1,
+          1024,
+          640
+        ]
+      },
+      "path": "parakeet_joint_decision.mlpackage"
+    },
+    "joint_decision_single_step": {
+      "inputs": {
+        "encoder_step": [
+          1,
+          1024,
+          1
+        ],
+        "decoder_step": [
+          1,
+          640,
+          1
+        ]
+      },
+      "outputs": {
+        "token_id": [
+          1,
+          1,
+          1
+        ],
+        "token_prob": [
+          1,
+          1,
+          1
+        ],
+        "duration": [
+          1,
+          1,
+          1
+        ]
+      },
+      "path": "parakeet_joint_decision_single_step.mlpackage"
+    }
+  }
+}