Upload 5 files

silero_vad_stateful_4bit.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5a6310ac1abb0ec7a06994d0c46e66c34afdf0fa23d88044eeb695e77011fcb0
+oid sha256:67d19ac70e21c7f01639678563b5cf6eb89bf4b15cbb7280b84dc204f30bfcbc
 size 243

silero_vad_stateful_4bit.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d945f4a9056d9eeebd6abc43d4d977f7a2da55d7fc6bb5993636ebc728fe371b
-size 529
+oid sha256:1d3c298161f5be74328c6c0253a12800e7a50b9a0b3c9342744eb78b3e4aef0a
+size 788

silero_vad_stateful_4bit.mlmodelc/metadata.json

@@ -1,15 +1,15 @@
 [
   {
-    "shortDescription" : "Stateful LSTM VAD with proper state management for iOS 17+",
+    "shortDescription" : "Production Silero VAD with LSTM states and full architecture",
     "metadataOutputVersion" : "3.0",
     "outputSchema" : [
       {
         "hasShapeFlexibility" : "0",
         "isOptional" : "0",
         "dataType" : "Float32",
-        "formattedType" : "MultiArray (Float32 1)",
-        "shortDescription" : "",
-        "shape" : "[1]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "shortDescription" : "Voice activity probability (0.0-1.0)",
+        "shape" : "[1, 1]",
         "name" : "vad_probability",
         "type" : "MultiArray"
       },
@@ -18,7 +18,7 @@
         "isOptional" : "0",
         "dataType" : "Float32",
         "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
-        "shortDescription" : "",
+        "shortDescription" : "Updated LSTM hidden state for next inference",
         "shape" : "[1, 1, 128]",
         "name" : "h_state_out",
         "type" : "MultiArray"
@@ -28,31 +28,44 @@
         "isOptional" : "0",
         "dataType" : "Float32",
         "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
-        "shortDescription" : "",
+        "shortDescription" : "Updated LSTM cell state for next inference",
         "shape" : "[1, 1, 128]",
         "name" : "c_state_out",
         "type" : "MultiArray"
       }
     ],
-    "version" : "1.0",
+    "version" : "3.0",
     "modelParameters" : [
 
     ],
-    "author" : "Clean Stateful VAD Model",
+    "author" : "Trained Stateful Silero VAD",
     "specificationVersion" : 8,
     "storagePrecision" : "Mixed (Float16, Palettized (4 bits))",
     "mlProgramOperationTypeHistogram" : {
-      "Ios16.sigmoid" : 1,
-      "Ios17.squeeze" : 4,
-      "Ios17.cast" : 6,
-      "Ios17.linear" : 4,
-      "Ios17.transpose" : 2,
-      "Ios16.relu" : 2,
-      "Ios16.constexprLutToDense" : 5,
+      "Ios17.square" : 6,
+      "Ios17.reshape" : 8,
+      "Ios16.reduceMean" : 5,
+      "Ios17.matmul" : 6,
+      "Ios17.transpose" : 9,
+      "Ios17.expandDims" : 2,
+      "Ios17.add" : 3,
+      "Ios16.sigmoid" : 5,
+      "Ios17.sliceByIndex" : 3,
+      "Ios17.squeeze" : 3,
+      "Ios17.gather" : 6,
+      "Ios17.layerNorm" : 1,
+      "Ios17.sqrt" : 3,
+      "Ios17.conv" : 5,
+      "Ios16.constexprLutToDense" : 8,
+      "Ios16.relu" : 9,
       "Ios17.lstm" : 1,
-      "Ios17.expandDims" : 3
+      "Ios17.linear" : 8,
+      "Stack" : 1,
+      "Ios17.concat" : 4,
+      "Ios17.cast" : 6,
+      "Ios17.mul" : 11
     },
-    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "computePrecision" : "Mixed (Float16, Float32, Int32, UInt16)",
     "stateSchema" : [
 
     ],
@@ -74,7 +87,7 @@
         "isOptional" : "0",
         "dataType" : "Float32",
         "formattedType" : "MultiArray (Float32 1 × 512)",
-        "shortDescription" : "",
+        "shortDescription" : "32ms audio chunk (512 samples at 16kHz)",
         "shape" : "[1, 512]",
         "name" : "audio_chunk",
         "type" : "MultiArray"
@@ -84,7 +97,7 @@
         "isOptional" : "0",
         "dataType" : "Float32",
         "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
-        "shortDescription" : "",
+        "shortDescription" : "LSTM hidden state from previous inference",
         "shape" : "[1, 1, 128]",
         "name" : "h_state_in",
         "type" : "MultiArray"
@@ -94,7 +107,7 @@
         "isOptional" : "0",
         "dataType" : "Float32",
         "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
-        "shortDescription" : "",
+        "shortDescription" : "LSTM cell state from previous inference",
         "shape" : "[1, 1, 128]",
         "name" : "c_state_in",
         "type" : "MultiArray"
@@ -102,10 +115,10 @@
     ],
     "userDefinedMetadata" : {
       "com.github.apple.coremltools.version" : "8.3.0",
-      "com.github.apple.coremltools.source" : "torch==2.5.0",
-      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.0"
     },
-    "generatedClassName" : "silero_vad_stateful_4bit_clean",
+    "generatedClassName" : "silero_vad_improved_stateful_4bit",
     "method" : "predict"
   }
 ]

silero_vad_stateful_4bit.mlmodelc/model.mil

@@ -2,59 +2,265 @@ program(1.0)
 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
 {
     func main<ios17>(tensor<fp32, [1, 512]> audio_chunk, tensor<fp32, [1, 1, 128]> c_state_in, tensor<fp32, [1, 1, 128]> h_state_in) {
+            tensor<int32, [2]> var_20_begin_0 = const()[name = tensor<string, []>("op_20_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_20_end_0 = const()[name = tensor<string, []>("op_20_end_0"), val = tensor<int32, [2]>([1, 256])];
+            tensor<bool, [2]> var_20_end_mask_0 = const()[name = tensor<string, []>("op_20_end_mask_0"), val = tensor<bool, [2]>([true, false])];
             tensor<string, []> audio_chunk_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_chunk_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
-            tensor<fp16, [128, 512]> encoder_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32896))), name = tensor<string, []>("encoder_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 512])];
-            tensor<fp16, [128]> encoder_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_0_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33024)))];
             tensor<fp16, [1, 512]> audio_chunk_to_fp16 = cast(dtype = audio_chunk_to_fp16_dtype_0, x = audio_chunk)[name = tensor<string, []>("cast_5")];
-            tensor<fp16, [1, 128]> linear_0_cast_fp16 = linear(bias = encoder_0_bias_to_fp16, weight = encoder_0_weight_to_fp16_palettized, x = audio_chunk_to_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
-            tensor<fp16, [1, 128]> input_3_cast_fp16 = relu(x = linear_0_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
-            tensor<fp16, [64, 128]> encoder_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37504))), name = tensor<string, []>("encoder_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([64, 128])];
-            tensor<fp16, [64]> encoder_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37632)))];
-            tensor<fp16, [1, 64]> linear_1_cast_fp16 = linear(bias = encoder_2_bias_to_fp16, weight = encoder_2_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
-            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 64]> input_5_cast_fp16 = expand_dims(axes = input_5_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 256]> var_20_cast_fp16 = slice_by_index(begin = var_20_begin_0, end = var_20_end_0, end_mask = var_20_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("op_20_cast_fp16")];
+            tensor<fp16, [256]> window_to_fp16 = const()[name = tensor<string, []>("window_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 256]> frame_1_cast_fp16 = mul(x = var_20_cast_fp16, y = window_to_fp16)[name = tensor<string, []>("frame_1_cast_fp16")];
+            tensor<bool, []> matmul_1_transpose_x_1 = const()[name = tensor<string, []>("matmul_1_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_1_transpose_y_1 = const()[name = tensor<string, []>("matmul_1_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 256]> cos_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(640))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33472))), name = tensor<string, []>("cos_0_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 256])];
+            tensor<fp16, [256, 1]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_1, transpose_y = matmul_1_transpose_y_1, x = cos_0_to_fp16_palettized, y = frame_1_cast_fp16)[name = tensor<string, []>("matmul_1_cast_fp16")];
+            tensor<bool, []> matmul_3_transpose_x_1 = const()[name = tensor<string, []>("matmul_3_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_3_transpose_y_1 = const()[name = tensor<string, []>("matmul_3_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 256]> sin_1_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33600))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66432))), name = tensor<string, []>("sin_1_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 256])];
+            tensor<fp16, [256, 1]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_1, transpose_y = matmul_3_transpose_y_1, x = sin_1_to_fp16_palettized, y = frame_1_cast_fp16)[name = tensor<string, []>("matmul_3_cast_fp16")];
+            tensor<fp16, []> mul_2_y_0_to_fp16 = const()[name = tensor<string, []>("mul_2_y_0_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
+            tensor<fp16, [256, 1]> mul_2_cast_fp16 = mul(x = matmul_3_cast_fp16, y = mul_2_y_0_to_fp16)[name = tensor<string, []>("mul_2_cast_fp16")];
+            tensor<int32, [2]> transpose_2_perm_0 = const()[name = tensor<string, []>("transpose_2_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, [2]> transpose_3_perm_0 = const()[name = tensor<string, []>("transpose_3_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<uint16, [129]> range_1d_2_to_uint16 = const()[name = tensor<string, []>("range_1d_2_to_uint16"), val = tensor<uint16, [129]>([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128])];
+            tensor<fp16, [1, 256]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = matmul_1_cast_fp16)[name = tensor<string, []>("transpose_8")];
+            tensor<fp16, [1, 129]> gather_0_cast_fp16_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_0_validate_indices_0, x = transpose_2_cast_fp16)[name = tensor<string, []>("gather_0_cast_fp16_cast_uint16")];
+            tensor<int32, []> gather_1_axis_0 = const()[name = tensor<string, []>("gather_1_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> gather_1_batch_dims_0 = const()[name = tensor<string, []>("gather_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_1_validate_indices_0 = const()[name = tensor<string, []>("gather_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 256]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = mul_2_cast_fp16)[name = tensor<string, []>("transpose_7")];
+            tensor<fp16, [1, 129]> gather_1_cast_fp16_cast_uint16 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_1_validate_indices_0, x = transpose_3_cast_fp16)[name = tensor<string, []>("gather_1_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 129]> square_0_cast_fp16 = square(x = gather_0_cast_fp16_cast_uint16)[name = tensor<string, []>("square_0_cast_fp16")];
+            tensor<fp16, [1, 129]> square_1_cast_fp16 = square(x = gather_1_cast_fp16_cast_uint16)[name = tensor<string, []>("square_1_cast_fp16")];
+            tensor<fp16, [1, 129]> add_2_cast_fp16 = add(x = square_0_cast_fp16, y = square_1_cast_fp16)[name = tensor<string, []>("add_2_cast_fp16")];
+            tensor<fp16, [1, 129]> sqrt_0_cast_fp16 = sqrt(x = add_2_cast_fp16)[name = tensor<string, []>("sqrt_0_cast_fp16")];
+            tensor<int32, [2]> var_25_begin_0 = const()[name = tensor<string, []>("op_25_begin_0"), val = tensor<int32, [2]>([0, 128])];
+            tensor<int32, [2]> var_25_end_0 = const()[name = tensor<string, []>("op_25_end_0"), val = tensor<int32, [2]>([1, 384])];
+            tensor<bool, [2]> var_25_end_mask_0 = const()[name = tensor<string, []>("op_25_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, 256]> var_25_cast_fp16 = slice_by_index(begin = var_25_begin_0, end = var_25_end_0, end_mask = var_25_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("op_25_cast_fp16")];
+            tensor<fp16, [1, 256]> frame_3_cast_fp16 = mul(x = var_25_cast_fp16, y = window_to_fp16)[name = tensor<string, []>("frame_3_cast_fp16")];
+            tensor<bool, []> matmul_6_transpose_x_1 = const()[name = tensor<string, []>("matmul_6_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_6_transpose_y_1 = const()[name = tensor<string, []>("matmul_6_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 1]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_1, transpose_y = matmul_6_transpose_y_1, x = cos_0_to_fp16_palettized, y = frame_3_cast_fp16)[name = tensor<string, []>("matmul_6_cast_fp16")];
+            tensor<bool, []> matmul_8_transpose_x_1 = const()[name = tensor<string, []>("matmul_8_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_8_transpose_y_1 = const()[name = tensor<string, []>("matmul_8_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 1]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_1, transpose_y = matmul_8_transpose_y_1, x = sin_1_to_fp16_palettized, y = frame_3_cast_fp16)[name = tensor<string, []>("matmul_8_cast_fp16")];
+            tensor<fp16, []> mul_4_y_0_to_fp16 = const()[name = tensor<string, []>("mul_4_y_0_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
+            tensor<fp16, [256, 1]> mul_4_cast_fp16 = mul(x = matmul_8_cast_fp16, y = mul_4_y_0_to_fp16)[name = tensor<string, []>("mul_4_cast_fp16")];
+            tensor<int32, [2]> transpose_6_perm_0 = const()[name = tensor<string, []>("transpose_6_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, [2]> transpose_7_perm_0 = const()[name = tensor<string, []>("transpose_7_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, []> gather_2_axis_0 = const()[name = tensor<string, []>("gather_2_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> gather_2_batch_dims_0 = const()[name = tensor<string, []>("gather_2_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_2_validate_indices_0 = const()[name = tensor<string, []>("gather_2_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 256]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = matmul_6_cast_fp16)[name = tensor<string, []>("transpose_6")];
+            tensor<fp16, [1, 129]> gather_2_cast_fp16_cast_uint16 = gather(axis = gather_2_axis_0, batch_dims = gather_2_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_2_validate_indices_0, x = transpose_6_cast_fp16)[name = tensor<string, []>("gather_2_cast_fp16_cast_uint16")];
+            tensor<int32, []> gather_3_axis_0 = const()[name = tensor<string, []>("gather_3_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> gather_3_batch_dims_0 = const()[name = tensor<string, []>("gather_3_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_3_validate_indices_0 = const()[name = tensor<string, []>("gather_3_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 256]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = mul_4_cast_fp16)[name = tensor<string, []>("transpose_5")];
+            tensor<fp16, [1, 129]> gather_3_cast_fp16_cast_uint16 = gather(axis = gather_3_axis_0, batch_dims = gather_3_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_3_validate_indices_0, x = transpose_7_cast_fp16)[name = tensor<string, []>("gather_3_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 129]> square_2_cast_fp16 = square(x = gather_2_cast_fp16_cast_uint16)[name = tensor<string, []>("square_2_cast_fp16")];
+            tensor<fp16, [1, 129]> square_3_cast_fp16 = square(x = gather_3_cast_fp16_cast_uint16)[name = tensor<string, []>("square_3_cast_fp16")];
+            tensor<fp16, [1, 129]> add_4_cast_fp16 = add(x = square_2_cast_fp16, y = square_3_cast_fp16)[name = tensor<string, []>("add_4_cast_fp16")];
+            tensor<fp16, [1, 129]> sqrt_1_cast_fp16 = sqrt(x = add_4_cast_fp16)[name = tensor<string, []>("sqrt_1_cast_fp16")];
+            tensor<int32, [2]> var_30_begin_0 = const()[name = tensor<string, []>("op_30_begin_0"), val = tensor<int32, [2]>([0, 256])];
+            tensor<int32, [2]> var_30_end_0 = const()[name = tensor<string, []>("op_30_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_30_end_mask_0 = const()[name = tensor<string, []>("op_30_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, 256]> var_30_cast_fp16 = slice_by_index(begin = var_30_begin_0, end = var_30_end_0, end_mask = var_30_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("op_30_cast_fp16")];
+            tensor<fp16, [1, 256]> frame_cast_fp16 = mul(x = var_30_cast_fp16, y = window_to_fp16)[name = tensor<string, []>("frame_cast_fp16")];
+            tensor<bool, []> matmul_11_transpose_x_1 = const()[name = tensor<string, []>("matmul_11_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_11_transpose_y_1 = const()[name = tensor<string, []>("matmul_11_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 1]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_1, transpose_y = matmul_11_transpose_y_1, x = cos_0_to_fp16_palettized, y = frame_cast_fp16)[name = tensor<string, []>("matmul_11_cast_fp16")];
+            tensor<bool, []> matmul_13_transpose_x_1 = const()[name = tensor<string, []>("matmul_13_transpose_x_1"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matmul_13_transpose_y_1 = const()[name = tensor<string, []>("matmul_13_transpose_y_1"), val = tensor<bool, []>(true)];
+            tensor<fp16, [256, 1]> matmul_13_cast_fp16 = matmul(transpose_x = matmul_13_transpose_x_1, transpose_y = matmul_13_transpose_y_1, x = sin_1_to_fp16_palettized, y = frame_cast_fp16)[name = tensor<string, []>("matmul_13_cast_fp16")];
+            tensor<fp16, []> mul_6_y_0_to_fp16 = const()[name = tensor<string, []>("mul_6_y_0_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
+            tensor<fp16, [256, 1]> mul_6_cast_fp16 = mul(x = matmul_13_cast_fp16, y = mul_6_y_0_to_fp16)[name = tensor<string, []>("mul_6_cast_fp16")];
+            tensor<int32, [2]> transpose_10_perm_0 = const()[name = tensor<string, []>("transpose_10_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, [2]> transpose_11_perm_0 = const()[name = tensor<string, []>("transpose_11_perm_0"), val = tensor<int32, [2]>([-1, 0])];
+            tensor<int32, []> gather_4_axis_0 = const()[name = tensor<string, []>("gather_4_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> gather_4_batch_dims_0 = const()[name = tensor<string, []>("gather_4_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_4_validate_indices_0 = const()[name = tensor<string, []>("gather_4_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 256]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = matmul_11_cast_fp16)[name = tensor<string, []>("transpose_4")];
+            tensor<fp16, [1, 129]> gather_4_cast_fp16_cast_uint16 = gather(axis = gather_4_axis_0, batch_dims = gather_4_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_4_validate_indices_0, x = transpose_10_cast_fp16)[name = tensor<string, []>("gather_4_cast_fp16_cast_uint16")];
+            tensor<int32, []> gather_5_axis_0 = const()[name = tensor<string, []>("gather_5_axis_0"), 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<fp16, [1, 256]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = mul_6_cast_fp16)[name = tensor<string, []>("transpose_3")];
+            tensor<fp16, [1, 129]> gather_5_cast_fp16_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = range_1d_2_to_uint16, validate_indices = gather_5_validate_indices_0, x = transpose_11_cast_fp16)[name = tensor<string, []>("gather_5_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 129]> square_4_cast_fp16 = square(x = gather_4_cast_fp16_cast_uint16)[name = tensor<string, []>("square_4_cast_fp16")];
+            tensor<fp16, [1, 129]> square_5_cast_fp16 = square(x = gather_5_cast_fp16_cast_uint16)[name = tensor<string, []>("square_5_cast_fp16")];
+            tensor<fp16, [1, 129]> add_6_cast_fp16 = add(x = square_4_cast_fp16, y = square_5_cast_fp16)[name = tensor<string, []>("add_6_cast_fp16")];
+            tensor<fp16, [1, 129]> sqrt_2_cast_fp16 = sqrt(x = add_6_cast_fp16)[name = tensor<string, []>("sqrt_2_cast_fp16")];
+            tensor<int32, []> input_1_axis_0 = const()[name = tensor<string, []>("input_1_axis_0"), val = tensor<int32, []>(2)];
+            tensor<fp16, [1, 129, 3]> input_1_cast_fp16 = stack(axis = input_1_axis_0, values = (sqrt_0_cast_fp16, sqrt_1_cast_fp16, sqrt_2_cast_fp16))[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<string, []> x_1_pad_type_0 = const()[name = tensor<string, []>("x_1_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> x_1_pad_0 = const()[name = tensor<string, []>("x_1_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> x_1_strides_0 = const()[name = tensor<string, []>("x_1_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> x_1_dilations_0 = const()[name = tensor<string, []>("x_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_1_groups_0 = const()[name = tensor<string, []>("x_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [128, 129, 3]> encoder_encoder_0_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66560))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91392))), name = tensor<string, []>("encoder_encoder_0_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 129, 3])];
+            tensor<fp16, [128]> encoder_encoder_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_0_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91520)))];
+            tensor<fp16, [1, 128, 3]> x_1_cast_fp16 = conv(bias = encoder_encoder_0_conv_bias_to_fp16, dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = encoder_encoder_0_conv_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
+            tensor<int32, [1]> reduce_mean_0_axes_0 = const()[name = tensor<string, []>("reduce_mean_0_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> reduce_mean_0_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_0_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 128, 1]> reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = x_1_cast_fp16)[name = tensor<string, []>("reduce_mean_0_cast_fp16")];
+            tensor<int32, []> concat_0_axis_0 = const()[name = tensor<string, []>("concat_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> concat_0_interleave_0 = const()[name = tensor<string, []>("concat_0_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 128, 1]> concat_0_cast_fp16 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = reduce_mean_0_cast_fp16)[name = tensor<string, []>("concat_0_cast_fp16")];
+            tensor<int32, [2]> var_65 = const()[name = tensor<string, []>("op_65"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_3_cast_fp16 = reshape(shape = var_65, x = concat_0_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [16, 128]> encoder_encoder_0_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_0_se_fc1_weight_to_fp16"), val = tensor<fp16, [16, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91840)))];
+            tensor<fp16, [16]> encoder_encoder_0_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_0_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96000)))];
+            tensor<fp16, [1, 16]> linear_0_cast_fp16 = linear(bias = encoder_encoder_0_se_fc1_bias_to_fp16, weight = encoder_encoder_0_se_fc1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [1, 16]> input_7_cast_fp16 = relu(x = linear_0_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [128, 16]> encoder_encoder_0_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_0_se_fc2_weight_to_fp16"), val = tensor<fp16, [128, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96128)))];
+            tensor<fp16, [128]> encoder_encoder_0_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_0_se_fc2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100288)))];
+            tensor<fp16, [1, 128]> linear_1_cast_fp16 = linear(bias = encoder_encoder_0_se_fc2_bias_to_fp16, weight = encoder_encoder_0_se_fc2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<fp16, [1, 128]> y_1_cast_fp16 = sigmoid(x = linear_1_cast_fp16)[name = tensor<string, []>("y_1_cast_fp16")];
+            tensor<int32, [3]> var_75 = const()[name = tensor<string, []>("op_75"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_3_cast_fp16 = reshape(shape = var_75, x = y_1_cast_fp16)[name = tensor<string, []>("y_3_cast_fp16")];
+            tensor<fp16, [1, 128, 3]> input_11_cast_fp16 = mul(x = x_1_cast_fp16, y = y_3_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 128, 3]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<string, []> x_3_pad_type_0 = const()[name = tensor<string, []>("x_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> x_3_pad_0 = const()[name = tensor<string, []>("x_3_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> x_3_strides_0 = const()[name = tensor<string, []>("x_3_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> x_3_dilations_0 = const()[name = tensor<string, []>("x_3_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_3_groups_0 = const()[name = tensor<string, []>("x_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [64, 128, 3]> encoder_encoder_1_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100608))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112960))), name = tensor<string, []>("encoder_encoder_1_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([64, 128, 3])];
+            tensor<fp16, [64]> encoder_encoder_1_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_1_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113088)))];
+            tensor<fp16, [1, 64, 3]> x_3_cast_fp16 = conv(bias = encoder_encoder_1_conv_bias_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = encoder_encoder_1_conv_weight_to_fp16_palettized, x = input_13_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<int32, [1]> reduce_mean_1_axes_0 = const()[name = tensor<string, []>("reduce_mean_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> reduce_mean_1_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_1_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 64, 1]> reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = x_3_cast_fp16)[name = tensor<string, []>("reduce_mean_1_cast_fp16")];
+            tensor<int32, []> concat_1_axis_0 = const()[name = tensor<string, []>("concat_1_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> concat_1_interleave_0 = const()[name = tensor<string, []>("concat_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 64, 1]> concat_1_cast_fp16 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = reduce_mean_1_cast_fp16)[name = tensor<string, []>("concat_1_cast_fp16")];
+            tensor<int32, [2]> var_94 = const()[name = tensor<string, []>("op_94"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_15_cast_fp16 = reshape(shape = var_94, x = concat_1_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_encoder_1_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_1_se_fc1_weight_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113280)))];
+            tensor<fp16, [8]> encoder_encoder_1_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_1_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 8]> linear_2_cast_fp16 = linear(bias = encoder_encoder_1_se_fc1_bias_to_fp16, weight = encoder_encoder_1_se_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, [1, 8]> input_19_cast_fp16 = relu(x = linear_2_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<fp16, [64, 8]> encoder_encoder_1_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_1_se_fc2_weight_to_fp16"), val = tensor<fp16, [64, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114368)))];
+            tensor<fp16, [64]> encoder_encoder_1_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_1_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115456)))];
+            tensor<fp16, [1, 64]> linear_3_cast_fp16 = linear(bias = encoder_encoder_1_se_fc2_bias_to_fp16, weight = encoder_encoder_1_se_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 64]> y_5_cast_fp16 = sigmoid(x = linear_3_cast_fp16)[name = tensor<string, []>("y_5_cast_fp16")];
+            tensor<int32, [3]> var_104 = const()[name = tensor<string, []>("op_104"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_7_cast_fp16 = reshape(shape = var_104, x = y_5_cast_fp16)[name = tensor<string, []>("y_7_cast_fp16")];
+            tensor<fp16, [1, 64, 3]> input_23_cast_fp16 = mul(x = x_3_cast_fp16, y = y_7_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
+            tensor<fp16, [1, 64, 3]> input_25_cast_fp16 = relu(x = input_23_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<string, []> x_5_pad_type_0 = const()[name = tensor<string, []>("x_5_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> x_5_pad_0 = const()[name = tensor<string, []>("x_5_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> x_5_strides_0 = const()[name = tensor<string, []>("x_5_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> x_5_dilations_0 = const()[name = tensor<string, []>("x_5_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_5_groups_0 = const()[name = tensor<string, []>("x_5_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [64, 64, 3]> encoder_encoder_2_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115648))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121856))), name = tensor<string, []>("encoder_encoder_2_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([64, 64, 3])];
+            tensor<fp16, [64]> encoder_encoder_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_2_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121984)))];
+            tensor<fp16, [1, 64, 3]> x_5_cast_fp16 = conv(bias = encoder_encoder_2_conv_bias_to_fp16, dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = encoder_encoder_2_conv_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
+            tensor<int32, [1]> reduce_mean_2_axes_0 = const()[name = tensor<string, []>("reduce_mean_2_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> reduce_mean_2_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_2_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 64, 1]> reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = x_5_cast_fp16)[name = tensor<string, []>("reduce_mean_2_cast_fp16")];
+            tensor<int32, []> concat_2_axis_0 = const()[name = tensor<string, []>("concat_2_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> concat_2_interleave_0 = const()[name = tensor<string, []>("concat_2_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 64, 1]> concat_2_cast_fp16 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = reduce_mean_2_cast_fp16)[name = tensor<string, []>("concat_2_cast_fp16")];
+            tensor<int32, [2]> var_123 = const()[name = tensor<string, []>("op_123"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_27_cast_fp16 = reshape(shape = var_123, x = concat_2_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [8, 64]> encoder_encoder_2_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_2_se_fc1_weight_to_fp16"), val = tensor<fp16, [8, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122176)))];
+            tensor<fp16, [8]> encoder_encoder_2_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_2_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0, 0x0p+0])];
+            tensor<fp16, [1, 8]> linear_4_cast_fp16 = linear(bias = encoder_encoder_2_se_fc1_bias_to_fp16, weight = encoder_encoder_2_se_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<fp16, [1, 8]> input_31_cast_fp16 = relu(x = linear_4_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<fp16, [64, 8]> encoder_encoder_2_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_2_se_fc2_weight_to_fp16"), val = tensor<fp16, [64, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(123264)))];
+            tensor<fp16, [64]> encoder_encoder_2_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_2_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124352)))];
+            tensor<fp16, [1, 64]> linear_5_cast_fp16 = linear(bias = encoder_encoder_2_se_fc2_bias_to_fp16, weight = encoder_encoder_2_se_fc2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<fp16, [1, 64]> y_9_cast_fp16 = sigmoid(x = linear_5_cast_fp16)[name = tensor<string, []>("y_9_cast_fp16")];
+            tensor<int32, [3]> var_133 = const()[name = tensor<string, []>("op_133"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_11_cast_fp16 = reshape(shape = var_133, x = y_9_cast_fp16)[name = tensor<string, []>("y_11_cast_fp16")];
+            tensor<fp16, [1, 64, 3]> input_35_cast_fp16 = mul(x = x_5_cast_fp16, y = y_11_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
+            tensor<fp16, [1, 64, 3]> input_37_cast_fp16 = relu(x = input_35_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<string, []> x_7_pad_type_0 = const()[name = tensor<string, []>("x_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> x_7_pad_0 = const()[name = tensor<string, []>("x_7_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> x_7_strides_0 = const()[name = tensor<string, []>("x_7_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> x_7_dilations_0 = const()[name = tensor<string, []>("x_7_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_7_groups_0 = const()[name = tensor<string, []>("x_7_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [128, 64, 3]> encoder_encoder_3_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124544))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136896))), name = tensor<string, []>("encoder_encoder_3_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 64, 3])];
+            tensor<fp16, [128]> encoder_encoder_3_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_3_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137024)))];
+            tensor<fp16, [1, 128, 3]> x_7_cast_fp16 = conv(bias = encoder_encoder_3_conv_bias_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = encoder_encoder_3_conv_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [1]> reduce_mean_3_axes_0 = const()[name = tensor<string, []>("reduce_mean_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> reduce_mean_3_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_3_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 128, 1]> reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = x_7_cast_fp16)[name = tensor<string, []>("reduce_mean_3_cast_fp16")];
+            tensor<int32, []> concat_3_axis_0 = const()[name = tensor<string, []>("concat_3_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> concat_3_interleave_0 = const()[name = tensor<string, []>("concat_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 128, 1]> concat_3_cast_fp16 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = reduce_mean_3_cast_fp16)[name = tensor<string, []>("concat_3_cast_fp16")];
+            tensor<int32, [2]> var_152 = const()[name = tensor<string, []>("op_152"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_39_cast_fp16 = reshape(shape = var_152, x = concat_3_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [16, 128]> encoder_encoder_3_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_3_se_fc1_weight_to_fp16"), val = tensor<fp16, [16, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137344)))];
+            tensor<fp16, [16]> encoder_encoder_3_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_3_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141504)))];
+            tensor<fp16, [1, 16]> linear_6_cast_fp16 = linear(bias = encoder_encoder_3_se_fc1_bias_to_fp16, weight = encoder_encoder_3_se_fc1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [1, 16]> input_43_cast_fp16 = relu(x = linear_6_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [128, 16]> encoder_encoder_3_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoder_3_se_fc2_weight_to_fp16"), val = tensor<fp16, [128, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141632)))];
+            tensor<fp16, [1, 128]> linear_7_cast_fp16 = linear(bias = encoder_encoder_0_se_fc2_bias_to_fp16, weight = encoder_encoder_3_se_fc2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<fp16, [1, 128]> y_13_cast_fp16 = sigmoid(x = linear_7_cast_fp16)[name = tensor<string, []>("y_13_cast_fp16")];
+            tensor<int32, [3]> var_162 = const()[name = tensor<string, []>("op_162"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_cast_fp16 = reshape(shape = var_162, x = y_13_cast_fp16)[name = tensor<string, []>("y_cast_fp16")];
+            tensor<fp16, [1, 128, 3]> input_47_cast_fp16 = mul(x = x_7_cast_fp16, y = y_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<fp16, [1, 128, 3]> x_9_cast_fp16 = relu(x = input_47_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<int32, [1]> x_11_axes_0 = const()[name = tensor<string, []>("x_11_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<bool, []> x_11_keep_dims_0 = const()[name = tensor<string, []>("x_11_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 128, 1]> x_11_cast_fp16 = reduce_mean(axes = x_11_axes_0, keep_dims = x_11_keep_dims_0, x = x_9_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<int32, [3]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [3]>([2, 0, 1])];
+            tensor<int32, [1]> input_49_batch_first_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_49_batch_first_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
             tensor<string, []> h_state_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_state_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
-            tensor<string, []> c_state_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_state_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
-            tensor<int32, [3]> input_5_batch_first_transpose_perm_0 = const()[name = tensor<string, []>("input_5_batch_first_transpose_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [1]> lstm_out_batch_first_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
             tensor<fp16, [1, 1, 128]> h_state_in_to_fp16 = cast(dtype = h_state_in_to_fp16_dtype_0, x = h_state_in)[name = tensor<string, []>("cast_4")];
-            tensor<fp16, [1, 128]> lstm_out_batch_first_lstm_h0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_h0_squeeze_axes_0, x = h_state_in_to_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_cast_fp16")];
-            tensor<int32, [1]> lstm_out_batch_first_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128]> input_49_batch_first_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_49_batch_first_lstm_h0_squeeze_axes_0, x = h_state_in_to_fp16)[name = tensor<string, []>("input_49_batch_first_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_49_batch_first_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_49_batch_first_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> c_state_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_state_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
             tensor<fp16, [1, 1, 128]> c_state_in_to_fp16 = cast(dtype = c_state_in_to_fp16_dtype_0, x = c_state_in)[name = tensor<string, []>("cast_3")];
-            tensor<fp16, [1, 128]> lstm_out_batch_first_lstm_c0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_c0_squeeze_axes_0, x = c_state_in_to_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_cast_fp16")];
-            tensor<string, []> lstm_out_batch_first_direction_0 = const()[name = tensor<string, []>("lstm_out_batch_first_direction_0"), val = tensor<string, []>("forward")];
-            tensor<bool, []> lstm_out_batch_first_output_sequence_0 = const()[name = tensor<string, []>("lstm_out_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
-            tensor<string, []> lstm_out_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
-            tensor<string, []> lstm_out_batch_first_cell_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<string, []> lstm_out_batch_first_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<fp16, [512, 64]> concat_1_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37824))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54272))), name = tensor<string, []>("concat_1_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 64])];
-            tensor<fp16, [512, 128]> concat_2_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54400))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87232))), name = tensor<string, []>("concat_2_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 128])];
-            tensor<fp16, [512]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87360)))];
-            tensor<fp16, [1, 1, 64]> input_5_batch_first_transpose_cast_fp16 = transpose(perm = input_5_batch_first_transpose_perm_0, x = input_5_cast_fp16)[name = tensor<string, []>("transpose_1")];
-            tensor<fp16, [1, 1, 128]> lstm_out_batch_first_cast_fp16_0, tensor<fp16, [1, 128]> lstm_out_batch_first_cast_fp16_1, tensor<fp16, [1, 128]> lstm_out_batch_first_cast_fp16_2 = lstm(activation = lstm_out_batch_first_activation_0, bias = concat_0_to_fp16, cell_activation = lstm_out_batch_first_cell_activation_0, direction = lstm_out_batch_first_direction_0, initial_c = lstm_out_batch_first_lstm_c0_squeeze_cast_fp16, initial_h = lstm_out_batch_first_lstm_h0_squeeze_cast_fp16, output_sequence = lstm_out_batch_first_output_sequence_0, recurrent_activation = lstm_out_batch_first_recurrent_activation_0, weight_hh = concat_2_to_fp16_palettized, weight_ih = concat_1_to_fp16_palettized, x = input_5_batch_first_transpose_cast_fp16)[name = tensor<string, []>("lstm_out_batch_first_cast_fp16")];
-            tensor<int32, [3]> lstm_out_perm_0 = const()[name = tensor<string, []>("lstm_out_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<int32, [1]> var_37_axes_0 = const()[name = tensor<string, []>("op_37_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp16, [1, 1, 128]> var_37_cast_fp16 = expand_dims(axes = var_37_axes_0, x = lstm_out_batch_first_cast_fp16_1)[name = tensor<string, []>("op_37_cast_fp16")];
-            tensor<string, []> var_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_37_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
-            tensor<int32, [1]> var_38_axes_0 = const()[name = tensor<string, []>("op_38_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp16, [1, 1, 128]> var_38_cast_fp16 = expand_dims(axes = var_38_axes_0, x = lstm_out_batch_first_cast_fp16_2)[name = tensor<string, []>("op_38_cast_fp16")];
-            tensor<string, []> var_38_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_38_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
-            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 1, 128]> lstm_out_cast_fp16 = transpose(perm = lstm_out_perm_0, x = lstm_out_batch_first_cast_fp16_0)[name = tensor<string, []>("transpose_0")];
-            tensor<fp16, [1, 128]> input_7_cast_fp16 = squeeze(axes = input_7_axes_0, x = lstm_out_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
-            tensor<fp16, [64, 128]> decoder_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88448))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92608))), name = tensor<string, []>("decoder_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([64, 128])];
-            tensor<fp16, [64]> decoder_0_bias_to_fp16 = const()[name = tensor<string, []>("decoder_0_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92736)))];
-            tensor<fp16, [1, 64]> linear_2_cast_fp16 = linear(bias = decoder_0_bias_to_fp16, weight = decoder_0_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
-            tensor<fp16, [1, 64]> input_11_cast_fp16 = relu(x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
-            tensor<fp16, [1, 64]> decoder_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_2_weight_to_fp16"), val = tensor<fp16, [1, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92928)))];
-            tensor<fp16, [1]> decoder_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_2_bias_to_fp16"), val = tensor<fp16, [1]>([-0x1.92p-4])];
-            tensor<fp16, [1, 1]> linear_3_cast_fp16 = linear(bias = decoder_2_bias_to_fp16, weight = decoder_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
-            tensor<fp16, [1, 1]> prob_cast_fp16 = sigmoid(x = linear_3_cast_fp16)[name = tensor<string, []>("prob_cast_fp16")];
-            tensor<int32, [1]> var_56_axes_0 = const()[name = tensor<string, []>("op_56_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1]> var_56_cast_fp16 = squeeze(axes = var_56_axes_0, x = prob_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
-            tensor<string, []> var_56_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_56_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
-            tensor<fp32, [1, 1, 128]> h_state_out = cast(dtype = var_37_cast_fp16_to_fp32_dtype_0, x = var_37_cast_fp16)[name = tensor<string, []>("cast_0")];
-            tensor<fp32, [1, 1, 128]> c_state_out = cast(dtype = var_38_cast_fp16_to_fp32_dtype_0, x = var_38_cast_fp16)[name = tensor<string, []>("cast_1")];
-            tensor<fp32, [1]> vad_probability = cast(dtype = var_56_cast_fp16_to_fp32_dtype_0, x = var_56_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp16, [1, 128]> input_49_batch_first_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_49_batch_first_lstm_c0_squeeze_axes_0, x = c_state_in_to_fp16)[name = tensor<string, []>("input_49_batch_first_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_49_batch_first_direction_0 = const()[name = tensor<string, []>("input_49_batch_first_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_49_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_49_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_49_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_49_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_49_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_49_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_49_batch_first_activation_0 = const()[name = tensor<string, []>("input_49_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [512, 128]> concat_5_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(145792))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178624))), name = tensor<string, []>("concat_5_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 128])];
+            tensor<fp16, [512, 128]> concat_6_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178752))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211584))), name = tensor<string, []>("concat_6_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 128])];
+            tensor<fp16, [512]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211712)))];
+            tensor<fp16, [1, 1, 128]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 128]> input_49_batch_first_cast_fp16_0, tensor<fp16, [1, 128]> input_49_batch_first_cast_fp16_1, tensor<fp16, [1, 128]> input_49_batch_first_cast_fp16_2 = lstm(activation = input_49_batch_first_activation_0, bias = concat_4_to_fp16, cell_activation = input_49_batch_first_cell_activation_0, direction = input_49_batch_first_direction_0, initial_c = input_49_batch_first_lstm_c0_squeeze_cast_fp16, initial_h = input_49_batch_first_lstm_h0_squeeze_cast_fp16, output_sequence = input_49_batch_first_output_sequence_0, recurrent_activation = input_49_batch_first_recurrent_activation_0, weight_hh = concat_6_to_fp16_palettized, weight_ih = concat_5_to_fp16_palettized, x = transpose_12_cast_fp16)[name = tensor<string, []>("input_49_batch_first_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> var_189_axes_0 = const()[name = tensor<string, []>("op_189_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> var_189_cast_fp16 = expand_dims(axes = var_189_axes_0, x = input_49_batch_first_cast_fp16_1)[name = tensor<string, []>("op_189_cast_fp16")];
+            tensor<string, []> var_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_189_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_190_axes_0 = const()[name = tensor<string, []>("op_190_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> var_190_cast_fp16 = expand_dims(axes = var_190_axes_0, x = input_49_batch_first_cast_fp16_2)[name = tensor<string, []>("op_190_cast_fp16")];
+            tensor<string, []> var_190_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_190_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_198_axes_0 = const()[name = tensor<string, []>("op_198_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> decoder_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("decoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212800)))];
+            tensor<fp16, []> var_172_to_fp16 = const()[name = tensor<string, []>("op_172_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 128]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = input_49_batch_first_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 128]> var_198_cast_fp16 = layer_norm(axes = var_198_axes_0, beta = encoder_encoder_0_se_fc2_bias_to_fp16, epsilon = var_172_to_fp16, gamma = decoder_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("op_198_cast_fp16")];
+            tensor<fp16, []> var_199_to_fp16 = const()[name = tensor<string, []>("op_199_to_fp16"), val = tensor<fp16, []>(0x1.334p-3)];
+            tensor<fp16, [1, 1, 128]> x_cast_fp16 = mul(x = var_198_cast_fp16, y = var_199_to_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<int32, [3]> input_51_perm_0 = const()[name = tensor<string, []>("input_51_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 128, 1]> input_51_cast_fp16 = transpose(perm = input_51_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 128, 1]> input_55_cast_fp16 = relu(x = input_51_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<string, []> input_pad_type_0 = const()[name = tensor<string, []>("input_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_strides_0 = const()[name = tensor<string, []>("input_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_pad_0 = const()[name = tensor<string, []>("input_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_dilations_0 = const()[name = tensor<string, []>("input_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_groups_0 = const()[name = tensor<string, []>("input_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [1, 128, 1]> decoder_conv_weight_to_fp16 = const()[name = tensor<string, []>("decoder_conv_weight_to_fp16"), val = tensor<fp16, [1, 128, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213120)))];
+            tensor<fp16, [1]> decoder_conv_bias_to_fp16 = const()[name = tensor<string, []>("decoder_conv_bias_to_fp16"), val = tensor<fp16, [1]>([-0x1.808p-6])];
+            tensor<fp16, [1, 1, 1]> input_cast_fp16 = conv(bias = decoder_conv_bias_to_fp16, dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = decoder_conv_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_211_cast_fp16 = sigmoid(x = input_cast_fp16)[name = tensor<string, []>("op_211_cast_fp16")];
+            tensor<int32, [1]> var_212_axes_0 = const()[name = tensor<string, []>("op_212_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_212_cast_fp16 = squeeze(axes = var_212_axes_0, x = var_211_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")];
+            tensor<string, []> var_212_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_212_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 128]> h_state_out = cast(dtype = var_189_cast_fp16_to_fp32_dtype_0, x = var_189_cast_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<fp32, [1, 1, 128]> c_state_out = cast(dtype = var_190_cast_fp16_to_fp32_dtype_0, x = var_190_cast_fp16)[name = tensor<string, []>("cast_1")];
+            tensor<fp32, [1, 1]> vad_probability = cast(dtype = var_212_cast_fp16_to_fp32_dtype_0, x = var_212_cast_fp16)[name = tensor<string, []>("cast_2")];
         } -> (vad_probability, h_state_out, c_state_out);
 }

silero_vad_stateful_4bit.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5f4767f28d716a87ffa0e446ae9dac70e1ca3708548e65691ae407071b56df8b
-size 93120
+oid sha256:2b948126693c78abcaf0a1dbaff2e7970b2481f5887c4c63cab1f73ded6a0537
+size 213440