Upload CoreML conversion of primeline/parakeet-primeline for FluidAudio (ANE)

Decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924034f7113ebd7f649e2f1e956fbd4b4401d1de331d3e35e86925d9e8f54cb9
+size 243

Decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1deb45675d84d6e0579486d693b5f49b166f6f20fba4c1b6bd768c227f9876f1
+size 554

Decoder.mlmodelc/model.mil

@@ -0,0 +1,73 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"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<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<fp16, [8193, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [8193, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_1_dtype_0 = const()[name = tensor<string, []>("cast_1_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_9")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = targets_to_int16)[name = tensor<string, []>("cast_8")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, 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 = cast_1, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("y_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16_cast_uint16")];
+            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<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            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<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            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<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            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<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            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<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10487168)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13764032)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17040896)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, 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_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_lstm_layer_0_cast_fp16")];
+            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<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            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<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            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<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17046080)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20322944)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23599808)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_4")];
+            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:d92ce68e4dcdb650b0446a390017e1a7470fd8c2b457585cd19c5909f8c3e5c5
+size 23604992

Encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6c3755b4576c49502e12c68b694a740ff3c8d782a9af1ab75ed87f06493f959
+size 243

Encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd074842dcd83dc331c3dab26c9d408bc82aee0597f4061e3e009010ddde7054
+size 478

Encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2594e5225447ac4d951685406c1f8329a7bf61a9df0ed015f9c6eab0d864fa66
+size 1185860032

JointDecisionv3.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aaad73f1755357028c9e42ee016d0c155b61fc9bc0cf2767465d863111fe880d
+size 243

JointDecisionv3.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:074a289abbd5ae0eef0e158302085bcbe0386e2e57412efbc64f7e5a1255d8f3
+size 591

JointDecisionv3.mlmodelc/model.mil

@@ -0,0 +1,69 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"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<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_step_to_fp16 = cast(dtype = encoder_step_to_fp16_dtype_0, x = encoder_step)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_step_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1312192)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_step_to_fp16 = cast(dtype = decoder_step_to_fp16_dtype_0, x = decoder_step)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_step_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [8198, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [8198, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132800)))];
+            tensor<fp16, [8198]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12626304)))];
+            tensor<fp16, [1, 1, 1, 8198]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            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<fp16, [1, 1, 1, 8193]> token_logits_cast_fp16 = slice_by_index(begin = token_logits_begin_0, end = token_logits_end_0, end_mask = token_logits_end_mask_0, x = linear_2_cast_fp16)[name = tensor<string, []>("token_logits_cast_fp16")];
+            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<fp16, [1, 1, 1, 5]> duration_logits_cast_fp16 = slice_by_index(begin = duration_logits_begin_0, end = duration_logits_end_0, end_mask = duration_logits_end_mask_0, x = linear_2_cast_fp16)[name = tensor<string, []>("duration_logits_cast_fp16")];
+            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_cast_fp16)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<int32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 1, 1, 8193]> token_probs_all_cast_fp16 = softmax(axis = var_49, x = token_logits_cast_fp16)[name = tensor<string, []>("token_probs_all_cast_fp16")];
+            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<string, []> var_58_to_int16_dtype_0 = const()[name = tensor<string, []>("op_58_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1, 1, 1]> var_58_to_int16 = cast(dtype = var_58_to_int16_dtype_0, x = var_58)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1, 1, 1]> var_61_cast_fp16_cast_int16 = gather_along_axis(axis = var_59, indices = var_58_to_int16, validate_indices = var_61_validate_indices_0, x = token_probs_all_cast_fp16)[name = tensor<string, []>("op_61_cast_fp16_cast_int16")];
+            tensor<int32, [1]> var_63_axes_0 = const()[name = tensor<string, []>("op_63_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1, 1]> var_63_cast_fp16 = squeeze(axes = var_63_axes_0, x = var_61_cast_fp16_cast_int16)[name = tensor<string, []>("op_63_cast_fp16")];
+            tensor<string, []> var_63_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            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_cast_fp16)[name = tensor<string, []>("op_66_cast_fp16")];
+            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_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_cast_int16_output_indices_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<fp16, [1, 1, 1, 64]> var_76_cast_fp16_cast_int16_0, tensor<uint16, [1, 1, 1, 64]> var_76_cast_fp16_cast_int16_1 = topk(ascending = var_76_ascending_0, axis = var_76_axis_0, k = var_72, output_indices_dtype = var_76_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_76_return_indices_0, sort = var_76_sort_0, x = token_logits_cast_fp16)[name = tensor<string, []>("op_76_cast_fp16_cast_int16")];
+            tensor<string, []> var_76_cast_fp16_cast_int16_1_to_int32_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_cast_int16_1_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_76_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits = cast(dtype = var_76_cast_fp16_0_to_fp32_dtype_0, x = var_76_cast_fp16_cast_int16_0)[name = tensor<string, []>("cast_0")];
+            tensor<int32, [1, 1, 1, 64]> top_k_ids = cast(dtype = var_76_cast_fp16_cast_int16_1_to_int32_dtype_0, x = var_76_cast_fp16_cast_int16_1)[name = tensor<string, []>("cast_1")];
+            tensor<fp32, [1, 1, 1]> token_prob = cast(dtype = var_63_cast_fp16_to_fp32_dtype_0, x = var_63_cast_fp16)[name = tensor<string, []>("cast_2")];
+        } -> (token_id, token_prob, duration, top_k_ids, top_k_logits);
+}

JointDecisionv3.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0c74f8b291468b11dd40c812a57df6bb880c084e8971ade72a45374f294b2d2
+size 12642764

Preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b54af4b97ed3eb0c970449fe7c83d86c4d7c0daa577a4b34fcd4b6696a32ca1b
+size 243

Preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2724a1551d8d3b514a47f792cb943f6cb7a6c7a88d836b0ee82ee1e725ec4f6a
+size 493

Preprocessor.mlmodelc/model.mil

@@ -0,0 +1,170 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"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<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_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_38_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> var_39_promoted_to_fp16 = const()[name = tensor<string, []>("op_39_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [1]> floor_div_0_to_fp16 = cast(dtype = var_38_to_fp16_dtype_0, x = floor_div_0)[name = tensor<string, []>("cast_22")];
+            tensor<fp16, [1]> seq_len_1_cast_fp16 = add(x = floor_div_0_to_fp16, y = var_39_promoted_to_fp16)[name = tensor<string, []>("seq_len_1_cast_fp16")];
+            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<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, ?]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_20")];
+            tensor<fp16, [1]> var_43_cast_fp16 = 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_to_fp16)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_44_cast_fp16 = expand_dims(axes = var_44_axes_0, x = var_43_cast_fp16)[name = tensor<string, []>("op_44_cast_fp16")];
+            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<fp16, [1, ?]> var_46_cast_fp16 = slice_by_index(begin = var_46_begin_0, end = var_46_end_0, end_mask = var_46_end_mask_0, x = audio_signal_to_fp16)[name = tensor<string, []>("op_46_cast_fp16")];
+            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<fp16, [1, ?]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, x = audio_signal_to_fp16)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<fp16, []> var_49_to_fp16 = const()[name = tensor<string, []>("op_49_to_fp16"), val = tensor<fp16, []>(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_50_cast_fp16 = mul(x = var_48_cast_fp16, y = var_49_to_fp16)[name = tensor<string, []>("op_50_cast_fp16")];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = sub(x = var_46_cast_fp16, y = var_50_cast_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<bool, []> input_1_interleave_0 = const()[name = tensor<string, []>("input_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = concat(axis = var_9, interleave = input_1_interleave_0, values = (var_44_cast_fp16, var_51_cast_fp16))[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_0x, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            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<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_1x, x = input_5_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            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<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = tensor<string, []>("expand_dims_4_cast_fp16")];
+            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<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = tensor<string, []>("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = 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_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_0_cast_fp16")];
+            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<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = tensor<string, []>("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = 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_to_fp16, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_1_cast_fp16")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = tensor<string, []>("stack_0_cast_fp16")];
+            tensor<fp16, []> var_17_promoted_to_fp16 = const()[name = tensor<string, []>("op_17_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_67_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_17_promoted_to_fp16)[name = tensor<string, []>("op_67_cast_fp16")];
+            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<fp16, [1, 257, ?]> var_69_cast_fp16 = reduce_sum(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_67_cast_fp16)[name = tensor<string, []>("op_69_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_9_cast_fp16 = identity(x = var_69_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            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<fp16, [1, 128, 257]> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526528)))];
+            tensor<fp16, [1, 128, ?]> x_11_cast_fp16 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = const_2_to_fp16, y = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<fp16, []> var_76_to_fp16 = const()[name = tensor<string, []>("op_76_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_77_cast_fp16 = add(x = x_11_cast_fp16, y = var_76_to_fp16)[name = tensor<string, []>("op_77_cast_fp16")];
+            tensor<fp32, []> x_13_epsilon_0 = const()[name = tensor<string, []>("x_13_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = log(epsilon = x_13_epsilon_0, x = var_77_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
+            tensor<int32, [3]> var_79_shape_cast_fp16 = shape(x = x_13_cast_fp16)[name = tensor<string, []>("op_79_shape_cast_fp16")];
+            tensor<int32, []> gather_4 = const()[name = tensor<string, []>("gather_4"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_5_axis_0 = const()[name = tensor<string, []>("gather_5_axis_0"), val = tensor<int32, []>(0)];
+            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<string, []> var_79_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_79_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<uint16, []> gather_5_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_5_indices_0_to_uint16"), val = tensor<uint16, []>(2)];
+            tensor<uint16, [3]> var_79_shape_cast_fp16_to_uint16 = cast(dtype = var_79_shape_cast_fp16_to_uint16_dtype_0, x = var_79_shape_cast_fp16)[name = tensor<string, []>("cast_19")];
+            tensor<uint16, []> gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = gather_5_indices_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_79_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_5_cast_uint16")];
+            tensor<string, []> gather_5_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_5_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            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, []> gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = tensor<string, []>("cast_18")];
+            tensor<int32, [?]> var_81 = range_1d(end = gather_5_cast_uint16_to_int32, 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_cast_uint16_to_int32))[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_cast_fp16)[name = tensor<string, []>("cast_21")];
+            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<fp16, []> var_24_to_fp16 = const()[name = tensor<string, []>("op_24_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 128, ?]> var_88_cast_fp16 = select(a = x_13_cast_fp16, b = var_24_to_fp16, cond = var_87)[name = tensor<string, []>("op_88_cast_fp16")];
+            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<fp16, [1, 128]> x_mean_numerator_cast_fp16 = reduce_sum(axes = x_mean_numerator_axes_0, keep_dims = x_mean_numerator_keep_dims_0, x = var_88_cast_fp16)[name = tensor<string, []>("x_mean_numerator_cast_fp16")];
+            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<string, []> cast_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [?, ?]> valid_mask_to_fp16 = cast(dtype = cast_3_to_fp16_dtype_0, x = valid_mask)[name = tensor<string, []>("cast_17")];
+            tensor<fp16, [?]> x_mean_denominator_cast_fp16 = reduce_sum(axes = x_mean_denominator_axes_0, keep_dims = x_mean_denominator_keep_dims_0, x = valid_mask_to_fp16)[name = tensor<string, []>("x_mean_denominator_cast_fp16")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [?, 1]> var_93_cast_fp16 = expand_dims(axes = var_93_axes_0, x = x_mean_denominator_cast_fp16)[name = tensor<string, []>("op_93_cast_fp16")];
+            tensor<fp16, [?, 128]> x_mean_cast_fp16 = real_div(x = x_mean_numerator_cast_fp16, y = var_93_cast_fp16)[name = tensor<string, []>("x_mean_cast_fp16")];
+            tensor<int32, [1]> var_96_axes_0 = const()[name = tensor<string, []>("op_96_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [?, 128, 1]> var_96_cast_fp16 = expand_dims(axes = var_96_axes_0, x = x_mean_cast_fp16)[name = tensor<string, []>("op_96_cast_fp16")];
+            tensor<fp16, [?, 128, ?]> var_97_cast_fp16 = sub(x = x_13_cast_fp16, y = var_96_cast_fp16)[name = tensor<string, []>("op_97_cast_fp16")];
+            tensor<fp16, [?, 128, ?]> var_98_cast_fp16 = select(a = var_97_cast_fp16, b = var_24_to_fp16, cond = var_87)[name = tensor<string, []>("op_98_cast_fp16")];
+            tensor<fp16, []> var_17_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_17_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [?, 128, ?]> var_99_cast_fp16 = pow(x = var_98_cast_fp16, y = var_17_promoted_1_to_fp16)[name = tensor<string, []>("op_99_cast_fp16")];
+            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<fp16, [?, 128]> var_101_cast_fp16 = reduce_sum(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_99_cast_fp16)[name = tensor<string, []>("op_101_cast_fp16")];
+            tensor<fp16, []> var_103_to_fp16 = const()[name = tensor<string, []>("op_103_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<fp16, [?, 1]> var_104_cast_fp16 = sub(x = var_93_cast_fp16, y = var_103_to_fp16)[name = tensor<string, []>("op_104_cast_fp16")];
+            tensor<fp16, [?, 128]> var_105_cast_fp16 = real_div(x = var_101_cast_fp16, y = var_104_cast_fp16)[name = tensor<string, []>("op_105_cast_fp16")];
+            tensor<fp16, [?, 128]> x_std_1_cast_fp16 = sqrt(x = var_105_cast_fp16)[name = tensor<string, []>("x_std_1_cast_fp16")];
+            tensor<fp16, []> var_25_to_fp16 = const()[name = tensor<string, []>("op_25_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [?, 128]> x_std_cast_fp16 = add(x = x_std_1_cast_fp16, y = var_25_to_fp16)[name = tensor<string, []>("x_std_cast_fp16")];
+            tensor<int32, [1]> var_110_axes_0 = const()[name = tensor<string, []>("op_110_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [?, 128, 1]> var_110_cast_fp16 = expand_dims(axes = var_110_axes_0, x = x_std_cast_fp16)[name = tensor<string, []>("op_110_cast_fp16")];
+            tensor<fp16, [?, 128, ?]> x_cast_fp16 = real_div(x = var_97_cast_fp16, y = var_110_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<int32, [3]> var_112_shape_cast_fp16 = shape(x = x_cast_fp16)[name = tensor<string, []>("op_112_shape_cast_fp16")];
+            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<string, []> var_112_shape_cast_fp16_to_uint16_dtype_0 = const()[name = tensor<string, []>("op_112_shape_cast_fp16_to_uint16_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<int32, []> gather_6_cast_uint16_axis_0 = const()[name = tensor<string, []>("gather_6_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<uint16, []> select_0_to_uint16 = const()[name = tensor<string, []>("select_0_to_uint16"), val = tensor<uint16, []>(2)];
+            tensor<uint16, [3]> var_112_shape_cast_fp16_to_uint16 = cast(dtype = var_112_shape_cast_fp16_to_uint16_dtype_0, x = var_112_shape_cast_fp16)[name = tensor<string, []>("cast_16")];
+            tensor<uint16, []> gather_6_cast_uint16_cast_uint16 = gather(axis = gather_6_cast_uint16_axis_0, batch_dims = gather_6_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_6_validate_indices_0, x = var_112_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_6_cast_uint16_cast_uint16")];
+            tensor<string, []> gather_6_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_6_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            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, []> gather_6_cast_uint16_to_int32 = cast(dtype = gather_6_cast_uint16_to_int32_dtype_0, x = gather_6_cast_uint16_cast_uint16)[name = tensor<string, []>("cast_15")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_6_cast_uint16_to_int32, start = const_5, step = const_6)[name = tensor<string, []>("mask_1")];
+            tensor<int32, []> gather_7_axis_0 = const()[name = tensor<string, []>("gather_7_axis_0"), val = tensor<int32, []>(0)];
+            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<uint16, []> gather_7_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_7_indices_0_to_uint16"), val = tensor<uint16, []>(0)];
+            tensor<uint16, []> gather_7_cast_uint16 = gather(axis = gather_7_axis_0, batch_dims = gather_7_batch_dims_0, indices = gather_7_indices_0_to_uint16, validate_indices = gather_7_validate_indices_0, x = var_112_shape_cast_fp16_to_uint16)[name = tensor<string, []>("gather_7_cast_uint16")];
+            tensor<string, []> gather_7_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_7_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            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, []> gather_7_cast_uint16_to_int32 = cast(dtype = gather_7_cast_uint16_to_int32_dtype_0, x = gather_7_cast_uint16)[name = tensor<string, []>("cast_14")];
+            tensor<int32, [2]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (gather_7_cast_uint16_to_int32, 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<fp16, [?, 128, ?]> processed_signal_cast_fp16 = select(a = var_24_to_fp16, b = x_cast_fp16, cond = var_119)[name = tensor<string, []>("processed_signal_cast_fp16")];
+            tensor<string, []> processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("processed_signal_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [?, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = tensor<string, []>("cast_13")];
+        } -> (mel, mel_length);
+}

Preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d374edc6e5cdb6e0429cfee7cc43b3a899150c3c3ebc24edcfb460046a3a9bf1
+size 592384

README.md

@@ -0,0 +1,78 @@
+---
+license: cc-by-4.0
+language:
+- de
+pipeline_tag: automatic-speech-recognition
+tags:
+- coreml
+- apple-neural-engine
+- fluidaudio
+- parakeet
+- german
+---
+
+# parakeet-primeline-de-coreml
+
+An unofficial CoreML conversion of [`primeline/parakeet-primeline`](https://huggingface.co/primeline/parakeet-primeline) — a German fine-tune of [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) — for offline speech-to-text on Apple Neural Engine via [FluidAudio](https://github.com/FluidInference/FluidAudio).
+
+All credit for the model itself goes to **primeline** (German fine-tuning) and **NVIDIA** (base FastConformer-TDT architecture and pretraining). This repository only provides a CoreML/ANE re-export of their weights; no retraining or fine-tuning was done here.
+
+## Why this exists
+
+`nvidia/parakeet-tdt-0.6b-v3` auto-detects language per utterance with no way to force it, so short German commands are sometimes misdecoded as English. primeline's fine-tune fixes this and improves German WER generally (see benchmarks below, reproduced from the [primeline/parakeet-primeline model card](https://huggingface.co/primeline/parakeet-primeline)). This repo makes that fine-tune usable on-device on Apple Silicon through FluidAudio's CoreML pipeline, instead of via NeMo/PyTorch on GPU.
+
+| Model | All (Avg) | Tuda-De | Multilingual LibriSpeech | Common Voice 19.0 |
+| :--- | :---: | :---: | :---: | :---: |
+| **primeline-parakeet** | **2.95** | **4.11** | **2.60** | **3.03** |
+| nvidia-parakeet-tdt-0.6b-v3 | 3.64 | 7.05 | 2.95 | 3.70 |
+
+## Conversion
+
+Converted with FluidInference's own [mobius](https://github.com/FluidInference/mobius) export tooling (`models/stt/parakeet-tdt-v3-0.6b/coreml/convert-parakeet.py`), pointed at primeline's `.nemo` checkpoint instead of the base NVIDIA model — no changes to the conversion script were needed, since primeline's fine-tune uses an unmodified tokenizer and architecture (verified byte-identical against the base model's vocabulary before conversion). Exported at FP16, `iOS17` deployment target, fixed 15-second audio window per FluidAudio's `v3` contract.
+
+Validated after conversion: exported component shapes match FluidAudio's expected `Preprocessor` / `Encoder` / `Decoder` / `JointDecisionv3` I/O contract exactly (including the `top_k_ids`/`top_k_logits` outputs FluidAudio's v3 decode path requires), and re-running the standard Tuda-De / Multilingual LibriSpeech / Common Voice 19.0 German benchmarks against the converted model landed close to primeline's published numbers, confirming the conversion didn't measurably degrade the model.
+
+`conversion_metadata.json` in this repo records the exact export settings.
+
+## Usage
+
+Drop-in for FluidAudio's `AsrModels.load(from:version:.v3, ...)` — same file layout as [`FluidInference/parakeet-tdt-0.6b-v3-coreml`](https://huggingface.co/FluidInference/parakeet-tdt-0.6b-v3-coreml), just German weights:
+
+```swift
+import FluidAudio
+
+let models = try await AsrModels.load(
+    from: modelsDirectory,   // a local directory literally named `parakeet-tdt-0.6b-v3`
+    version: .v3,
+    encoderPrecision: .int8,  // filename mapping only — this encoder is FP16, see note below
+    encoderComputeUnits: .cpuAndNeuralEngine
+)
+let manager = AsrManager(config: .default)
+try await manager.loadModels(models)
+```
+
+**Note on the directory name:** FluidAudio's loader derives the model folder name from its internal `Repo` enum and expects it to be exactly `parakeet-tdt-0.6b-v3`, regardless of which weights are inside. Place these files in a folder with that exact name.
+
+**Note on precision:** `encoderPrecision: .int8` selects the *filename* FluidAudio looks for (`Encoder.mlmodelc`), not the actual on-disk precision — this export is FP16, not int8-quantized, so it will be larger (~1.2 GB) than FluidAudio's official English encoder (~425 MB, which is quantized). Int8 quantization was not applied here to avoid an extra accuracy-validation pass; feel free to quantize further at your own risk.
+
+**Set `DownloadUtils.enforceOffline = true`** before loading if you're pointing at a private local path — otherwise a load failure will cause FluidAudio to delete the folder and attempt to re-download the *official English* v3 from HuggingFace.
+
+## Files
+
+| File | Description |
+|---|---|
+| `Preprocessor.mlmodelc` | Mel-spectrogram frontend |
+| `Encoder.mlmodelc` | FastConformer encoder (FP16) |
+| `Decoder.mlmodelc` | TDT prediction network |
+| `JointDecisionv3.mlmodelc` | Joint + greedy decision head (incl. top-K for FluidAudio's language-script filter) |
+| `parakeet_vocab.json` | SentencePiece vocabulary (identical to base v3's, 8192 tokens) |
+| `conversion_metadata.json` | Export settings and component I/O shapes from the conversion run |
+
+## License
+
+CC-BY-4.0, inherited from primeline's fine-tune (itself CC-BY-4.0, derived from NVIDIA's CC-BY-4.0 base model). Attribution:
+
+- Base architecture and pretraining: [NVIDIA `parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3)
+- German fine-tuning: [primeline `parakeet-primeline`](https://huggingface.co/primeline/parakeet-primeline)
+- CoreML conversion tooling: [FluidInference `mobius`](https://github.com/FluidInference/mobius)
+- Runtime SDK this is built for: [FluidInference `FluidAudio`](https://github.com/FluidInference/FluidAudio)

conversion_metadata.json

@@ -0,0 +1,212 @@
+{
+  "model_id": "primeline/parakeet-primeline",
+  "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": "file",
+    "path": "/Users/valentinweyer/Dropbox/Valentin/Projekte/AI/german-stt/convert/ckpt/2_95_WER.nemo"
+  },
+  "coreml": {
+    "compute_units": "CPU_ONLY",
+    "compute_precision": "FLOAT16"
+  },
+  "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"
+    }
+  }
+}