Duplicate from FluidInference/silero-vad-coreml

Co-authored-by: Alex Weng <alexwengg@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,51 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+test/ambient1.mp3 filter=lfs diff=lfs merge=lfs -text
+test/ambient2.mp3 filter=lfs diff=lfs merge=lfs -text
+test/ambient3.mp3 filter=lfs diff=lfs merge=lfs -text
+test/ambient4.mp3 filter=lfs diff=lfs merge=lfs -text
+test/ambient5.mp3 filter=lfs diff=lfs merge=lfs -text
+test/human1.mp3 filter=lfs diff=lfs merge=lfs -text
+test/human3.mp3 filter=lfs diff=lfs merge=lfs -text
+test/human5.mp3 filter=lfs diff=lfs merge=lfs -text
+silero_encoder.mlmodelc/model.espresso.weights filter=lfs diff=lfs merge=lfs -text
+silero_rnn_decoder.mlmodelc/model.espresso.weights filter=lfs diff=lfs merge=lfs -text
+silero_stft.mlmodelc/model.espresso.weights filter=lfs diff=lfs merge=lfs -text
+comparison.png filter=lfs diff=lfs merge=lfs -text
+speed.png filter=lfs diff=lfs merge=lfs -text
+graphs/yc_256ms_comparison_20250915_205721_2c04b81.png filter=lfs diff=lfs merge=lfs -text
+graphs/yc_performance_20250915_205721_2c04b81.png filter=lfs diff=lfs merge=lfs -text
+graphs/yc_standard_comparison_20250915_205721_2c04b81.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,94 @@
+---
+license: mit
+tags:
+- audio
+- voice-activity-detection
+- coreml
+- silero
+- speech
+- ios
+- macos
+- swift
+library_name: coreml
+pipeline_tag: voice-activity-detection
+datasets:
+- alexwengg/musan_mini50
+- alexwengg/musan_mini100
+metrics:
+- accuracy
+- f1
+language:
+- en
+base_model:
+- onnx-community/silero-vad
+---
+
+
+# **<span style="color:#5DAF8D">🧃 CoreML Silero VAD </span>**
+[![Discord](https://img.shields.io/badge/Discord-Join%20Chat-7289da.svg)](https://discord.gg/WNsvaCtmDe)
+[![GitHub Repo stars](https://img.shields.io/github/stars/FluidInference/FluidAudio?style=flat&logo=github)](https://github.com/FluidInference/FluidAudio)
+
+A CoreML implementation of the Silero Voice Activity
+Detection (VAD) model, optimized for Apple platforms
+(iOS/macOS). This repository contains pre-converted
+CoreML models ready for use in Swift applications.
+
+See FluidAudio Repo link at the top for more information
+
+## Model Description
+
+**Developed by:** Silero Team (original), converted by
+FluidAudio
+
+**Model type:** Voice Activity Detection
+
+**License:** MIT
+
+**Parent Model:**
+[silero-vad](https://github.com/snakers4/silero-vad)
+
+
+This is how the model performs against the silero-vad v6.0.0 basline Pytorch JIT version 
+
+![graphs/yc_standard_comparison_20250915_205721_2c04b81.png](graphs/yc_standard_comparison_20250915_205721_2c04b81.png)
+![graphs/yc_256ms_comparison_20250915_205721_2c04b81.png](graphs/yc_256ms_comparison_20250915_205721_2c04b81.png)
+
+Note that we tested the quantized versions, as the model is already tiny, theres no performance imporvement at all. 
+
+
+This is how the different models compare in terms of speed, the 256s takes in 8 chunks of 32ms and processes it in batches so its much faster
+![graphs/yc_performance_20250915_205721_2c04b81.png](graphs/yc_performance_20250915_205721_2c04b81.png)
+
+
+Conversion code is available here: [FluidInference/mobius](https://github.com/FluidInference/mobius)
+
+## Intended Use
+
+### Primary Use Cases
+- Real-time voice activity detection in iOS/macOS
+applications
+- Speech preprocessing for ASR systems
+- Audio segmentation and filtering
+
+## How to Use
+
+Citation
+
+@misc{silero-vad-coreml,
+  title={CoreML Silero VAD},
+  author={FluidAudio Team},
+  year={2024},
+
+url={https://huggingface.co/alexwengg/coreml-silero-vad}
+}
+
+@misc{silero-vad,
+  title={Silero VAD},
+  author={Silero Team},
+  year={2021},
+  url={https://github.com/snakers4/silero-vad}
+}
+
+
+- GitHub: https://github.com/FluidAudio/FluidAudioSwift
+

config.json

@@ -0,0 +1 @@
+{}

silero-vad-unified-256ms-v6.0.0.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30945d54e32c3f15ec35dc6ee32128a27a6cdc03b0a12ffab04434069c49dfb5
+size 243

silero-vad-unified-256ms-v6.0.0.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c3063bd09ba71c26ede0308d7c33591d0770e971a3fcc603ccad7ba1e8fb88d
+size 625

silero-vad-unified-256ms-v6.0.0.mlmodelc/metadata.json

@@ -0,0 +1,120 @@
+[
+  {
+    "shortDescription" : "Silero VAD Unified Model 256ms (STFT + Encoder + Decoder) with noisy-OR aggregation",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "vad_output",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "new_hidden_state",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "new_cell_state",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "6.0.0",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Infernece + Silero Team",
+    "specificationVersion" : 6,
+    "storagePrecision" : "Mixed (Float16, Float32)",
+    "mlProgramOperationTypeHistogram" : {
+      "Concat" : 9,
+      "Lstm" : 8,
+      "SliceByIndex" : 41,
+      "Clip" : 32,
+      "Pow" : 16,
+      "Transpose" : 16,
+      "Sub" : 2,
+      "Relu" : 40,
+      "Squeeze" : 18,
+      "Cast" : 54,
+      "Sigmoid" : 8,
+      "Add" : 16,
+      "ExpandDims" : 26,
+      "Sqrt" : 8,
+      "Mul" : 7,
+      "Conv" : 48,
+      "Pad" : 8
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "12.0",
+      "tvOS" : "15.0",
+      "visionOS" : "1.0",
+      "watchOS" : "8.0",
+      "iOS" : "15.0",
+      "macCatalyst" : "15.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 4160)",
+        "shortDescription" : "",
+        "shape" : "[1, 4160]",
+        "name" : "audio_input",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "hidden_state",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "cell_state",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-09-15",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "silero_vad_unified_256ms_v6_0_0",
+    "method" : "predict"
+  }
+]

silero-vad-unified-256ms-v6.0.0.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:853cf34740d3f5061f977ebe2976f7c921b064261c9c4753b3a1196f2dba42b4
+size 882304

silero-vad-unified-256ms-v6.0.0.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a996535cd1f1ac5f92bab1a66edb56bf054ef48c8742338b07ec5b892e2613c2
+size 157925

silero-vad-unified-256ms-v6.0.0.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:853cf34740d3f5061f977ebe2976f7c921b064261c9c4753b3a1196f2dba42b4
+size 882304

silero-vad-unified-256ms-v6.0.0.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "413A9C09-1AD3-4AA9-B5E2-5C5BA0FB1FCD": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "DFAFD1E0-B2F3-4E3C-80EC-C2F0FC9E061D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "413A9C09-1AD3-4AA9-B5E2-5C5BA0FB1FCD"
+}

silero-vad-unified-v6.0.0.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2141be60ea0adf7acb1232fbcfaffb2be308ae02e6672d3762aedf36611ea9fd
+size 243

silero-vad-unified-v6.0.0.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f460dcdf796b19c04bc38ab6e69601831f634e5e74d499487f0c8fe17ca12f0f
+size 593

silero-vad-unified-v6.0.0.mlmodelc/metadata.json

@@ -0,0 +1,117 @@
+[
+  {
+    "shortDescription" : "Silero VAD Unified Model (STFT + Encoder + Decoder)",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "vad_output",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "new_hidden_state",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "new_cell_state",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "6.0.0",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Infernece + Silero Team",
+    "specificationVersion" : 6,
+    "storagePrecision" : "Mixed (Float16, Float32)",
+    "mlProgramOperationTypeHistogram" : {
+      "Lstm" : 1,
+      "SliceByIndex" : 2,
+      "Clip" : 4,
+      "Transpose" : 2,
+      "Pow" : 2,
+      "Relu" : 5,
+      "Squeeze" : 4,
+      "Cast" : 12,
+      "Sigmoid" : 1,
+      "Add" : 2,
+      "ExpandDims" : 5,
+      "Sqrt" : 1,
+      "Conv" : 6,
+      "Pad" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "12.0",
+      "tvOS" : "15.0",
+      "visionOS" : "1.0",
+      "watchOS" : "8.0",
+      "iOS" : "15.0",
+      "macCatalyst" : "15.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 576)",
+        "shortDescription" : "",
+        "shape" : "[1, 576]",
+        "name" : "audio_input",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "hidden_state",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "cell_state",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-09-15",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "silero_vad_unified_v6_0_0",
+    "method" : "predict"
+  }
+]

silero-vad-unified-v6.0.0.mlmodelc/model.mil

@@ -0,0 +1,143 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios15>(tensor<fp32, [1, 576]> audio_input, tensor<fp32, [1, 128]> cell_state, tensor<fp32, [1, 128]> hidden_state) {
+            tensor<int32, [4]> x_1_pad_0 = const()[name = tensor<string, []>("x_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
+            tensor<string, []> x_1_mode_0 = const()[name = tensor<string, []>("x_1_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<string, []> audio_input_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_input_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 576]> audio_input_to_fp16 = cast(dtype = audio_input_to_fp16_dtype_0, x = audio_input)[name = tensor<string, []>("cast_11")];
+            tensor<fp16, [1, 640]> x_1_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = x_1_mode_0, pad = x_1_pad_0, x = audio_input_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
+            tensor<int32, [1]> x_3_axes_0 = const()[name = tensor<string, []>("x_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 640]> x_3_cast_fp16 = expand_dims(axes = x_3_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<string, []> stft_out_pad_type_0 = const()[name = tensor<string, []>("stft_out_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> stft_out_strides_0 = const()[name = tensor<string, []>("stft_out_strides_0"), val = tensor<int32, [1]>([128])];
+            tensor<int32, [2]> stft_out_pad_0 = const()[name = tensor<string, []>("stft_out_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> stft_out_dilations_0 = const()[name = tensor<string, []>("stft_out_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> stft_out_groups_0 = const()[name = tensor<string, []>("stft_out_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [258, 1, 256]> stft_forward_basis_to_fp16 = const()[name = tensor<string, []>("stft_forward_basis_to_fp16"), val = tensor<fp16, [258, 1, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [1, 258, 4]> stft_out_cast_fp16 = conv(dilations = stft_out_dilations_0, groups = stft_out_groups_0, pad = stft_out_pad_0, pad_type = stft_out_pad_type_0, strides = stft_out_strides_0, weight = stft_forward_basis_to_fp16, x = x_3_cast_fp16)[name = tensor<string, []>("stft_out_cast_fp16")];
+            tensor<int32, [3]> var_28_begin_0 = const()[name = tensor<string, []>("op_28_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> var_28_end_0 = const()[name = tensor<string, []>("op_28_end_0"), val = tensor<int32, [3]>([1, 129, 4])];
+            tensor<bool, [3]> var_28_end_mask_0 = const()[name = tensor<string, []>("op_28_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<fp16, [1, 129, 4]> var_28_cast_fp16 = slice_by_index(begin = var_28_begin_0, end = var_28_end_0, end_mask = var_28_end_mask_0, x = stft_out_cast_fp16)[name = tensor<string, []>("op_28_cast_fp16")];
+            tensor<int32, [3]> var_31_begin_0 = const()[name = tensor<string, []>("op_31_begin_0"), val = tensor<int32, [3]>([0, 129, 0])];
+            tensor<int32, [3]> var_31_end_0 = const()[name = tensor<string, []>("op_31_end_0"), val = tensor<int32, [3]>([1, 258, 4])];
+            tensor<bool, [3]> var_31_end_mask_0 = const()[name = tensor<string, []>("op_31_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 129, 4]> var_31_cast_fp16 = slice_by_index(begin = var_31_begin_0, end = var_31_end_0, end_mask = var_31_end_mask_0, x = stft_out_cast_fp16)[name = tensor<string, []>("op_31_cast_fp16")];
+            tensor<fp16, []> var_7_promoted_to_fp16 = const()[name = tensor<string, []>("op_7_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129, 4]> var_33_cast_fp16 = pow(x = var_28_cast_fp16, y = var_7_promoted_to_fp16)[name = tensor<string, []>("op_33_cast_fp16")];
+            tensor<fp16, []> var_7_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_7_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129, 4]> var_34_cast_fp16 = pow(x = var_31_cast_fp16, y = var_7_promoted_1_to_fp16)[name = tensor<string, []>("op_34_cast_fp16")];
+            tensor<fp16, [1, 129, 4]> var_35_cast_fp16 = add(x = var_33_cast_fp16, y = var_34_cast_fp16)[name = tensor<string, []>("op_35_cast_fp16")];
+            tensor<fp16, []> var_36_to_fp16 = const()[name = tensor<string, []>("op_36_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 129, 4]> var_37_cast_fp16 = add(x = var_35_cast_fp16, y = var_36_to_fp16)[name = tensor<string, []>("op_37_cast_fp16")];
+            tensor<fp16, [1, 129, 4]> input_1_cast_fp16 = sqrt(x = var_37_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_3_strides_0 = const()[name = tensor<string, []>("input_3_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_3_dilations_0 = const()[name = tensor<string, []>("input_3_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_3_groups_0 = const()[name = tensor<string, []>("input_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [128, 129, 3]> encoder_layers_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_weight_to_fp16"), val = tensor<fp16, [128, 129, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132224)))];
+            tensor<fp16, [128]> encoder_layers_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_0_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231360)))];
+            tensor<fp16, [1, 128, 4]> input_3_cast_fp16 = conv(bias = encoder_layers_0_bias_to_fp16, dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = encoder_layers_0_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [1, 128, 4]> x_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")];
+            tensor<fp16, []> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, []> var_40_to_fp16 = const()[name = tensor<string, []>("op_40_to_fp16"), val = tensor<fp16, []>(0x1.388p+13)];
+            tensor<fp16, [1, 128, 4]> clip_0_cast_fp16 = clip(alpha = const_1_to_fp16, beta = var_40_to_fp16, x = x_5_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<string, []> input_7_pad_type_0 = const()[name = tensor<string, []>("input_7_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_7_strides_0 = const()[name = tensor<string, []>("input_7_strides_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1]> input_7_dilations_0 = const()[name = tensor<string, []>("input_7_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_7_groups_0 = const()[name = tensor<string, []>("input_7_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [64, 128, 3]> encoder_layers_2_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_weight_to_fp16"), val = tensor<fp16, [64, 128, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231680)))];
+            tensor<fp16, [64]> encoder_layers_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280896)))];
+            tensor<fp16, [1, 64, 2]> input_7_cast_fp16 = conv(bias = encoder_layers_2_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = encoder_layers_2_weight_to_fp16, x = clip_0_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1, 64, 2]> x_7_cast_fp16 = relu(x = input_7_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 64, 2]> clip_1_cast_fp16 = clip(alpha = const_2_to_fp16, beta = var_40_to_fp16, x = x_7_cast_fp16)[name = tensor<string, []>("clip_1_cast_fp16")];
+            tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_11_strides_0 = const()[name = tensor<string, []>("input_11_strides_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1]> input_11_dilations_0 = const()[name = tensor<string, []>("input_11_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_11_groups_0 = const()[name = tensor<string, []>("input_11_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [64, 64, 3]> encoder_layers_4_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_weight_to_fp16"), val = tensor<fp16, [64, 64, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281088)))];
+            tensor<fp16, [64]> encoder_layers_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_4_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305728)))];
+            tensor<fp16, [1, 64, 1]> input_11_cast_fp16 = conv(bias = encoder_layers_4_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = encoder_layers_4_weight_to_fp16, x = clip_1_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 64, 1]> x_9_cast_fp16 = relu(x = input_11_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
+            tensor<fp16, []> const_3_to_fp16 = const()[name = tensor<string, []>("const_3_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 64, 1]> clip_2_cast_fp16 = clip(alpha = const_3_to_fp16, beta = var_40_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("clip_2_cast_fp16")];
+            tensor<string, []> input_15_pad_type_0 = const()[name = tensor<string, []>("input_15_pad_type_0"), val = tensor<string, []>("custom")];
+            tensor<int32, [2]> input_15_pad_0 = const()[name = tensor<string, []>("input_15_pad_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [1]> input_15_strides_0 = const()[name = tensor<string, []>("input_15_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> input_15_dilations_0 = const()[name = tensor<string, []>("input_15_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_15_groups_0 = const()[name = tensor<string, []>("input_15_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [128, 64, 3]> encoder_layers_6_weight_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_weight_to_fp16"), val = tensor<fp16, [128, 64, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305920)))];
+            tensor<fp16, [128]> encoder_layers_6_bias_to_fp16 = const()[name = tensor<string, []>("encoder_layers_6_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355136)))];
+            tensor<fp16, [1, 128, 1]> input_15_cast_fp16 = conv(bias = encoder_layers_6_bias_to_fp16, dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = encoder_layers_6_weight_to_fp16, x = clip_2_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, [1, 128, 1]> x_11_cast_fp16 = relu(x = input_15_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<fp16, []> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, []>(-inf)];
+            tensor<fp16, [1, 128, 1]> clip_3_cast_fp16 = clip(alpha = const_4_to_fp16, beta = var_40_to_fp16, x = x_11_cast_fp16)[name = tensor<string, []>("clip_3_cast_fp16")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([2, 0, 1])];
+            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<int32, [1]> hx_1_axes_0 = const()[name = tensor<string, []>("hx_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> hidden_state_to_fp16_dtype_0 = const()[name = tensor<string, []>("hidden_state_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128]> hidden_state_to_fp16 = cast(dtype = hidden_state_to_fp16_dtype_0, x = hidden_state)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1, 1, 128]> hx_1_cast_fp16 = expand_dims(axes = hx_1_axes_0, x = hidden_state_to_fp16)[name = tensor<string, []>("hx_1_cast_fp16")];
+            tensor<int32, [1]> hx_axes_0 = const()[name = tensor<string, []>("hx_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> cell_state_to_fp16_dtype_0 = const()[name = tensor<string, []>("cell_state_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128]> cell_state_to_fp16 = cast(dtype = cell_state_to_fp16_dtype_0, x = cell_state)[name = tensor<string, []>("cast_8")];
+            tensor<fp16, [1, 1, 128]> hx_cast_fp16 = expand_dims(axes = hx_axes_0, x = cell_state_to_fp16)[name = tensor<string, []>("hx_cast_fp16")];
+            tensor<fp32, [512]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355456)))];
+            tensor<fp32, [512, 128]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357568)))];
+            tensor<fp32, [512, 128]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(619776)))];
+            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, 128]> lstm_out_batch_first_lstm_h0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_h0_squeeze_axes_0, x = hx_1_cast_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_cast_fp16")];
+            tensor<string, []> lstm_out_batch_first_lstm_h0_squeeze_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            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]> lstm_out_batch_first_lstm_c0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_c0_squeeze_axes_0, x = hx_cast_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> lstm_out_batch_first_lstm_c0_squeeze_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            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<fp32, [1, 128]> lstm_out_batch_first_lstm_c0_squeeze_cast_fp16_to_fp32 = cast(dtype = lstm_out_batch_first_lstm_c0_squeeze_cast_fp16_to_fp32_dtype_0, x = lstm_out_batch_first_lstm_c0_squeeze_cast_fp16)[name = tensor<string, []>("cast_6")];
+            tensor<fp32, [1, 128]> lstm_out_batch_first_lstm_h0_squeeze_cast_fp16_to_fp32 = cast(dtype = lstm_out_batch_first_lstm_h0_squeeze_cast_fp16_to_fp32_dtype_0, x = lstm_out_batch_first_lstm_h0_squeeze_cast_fp16)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 128]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = clip_3_cast_fp16)[name = tensor<string, []>("transpose_3")];
+            tensor<fp32, [1, 1, 128]> transpose_0_cast_fp16_to_fp32 = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_10")];
+            tensor<fp32, [1, 1, 128]> lstm_out_batch_first_0, tensor<fp32, [1, 128]> lstm_out_batch_first_1, tensor<fp32, [1, 128]> lstm_out_batch_first_2 = lstm(activation = lstm_out_batch_first_activation_0, bias = concat_0, 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_to_fp32, initial_h = lstm_out_batch_first_lstm_h0_squeeze_cast_fp16_to_fp32, output_sequence = lstm_out_batch_first_output_sequence_0, recurrent_activation = lstm_out_batch_first_recurrent_activation_0, weight_hh = concat_2, weight_ih = concat_1, x = transpose_0_cast_fp16_to_fp32)[name = tensor<string, []>("lstm_out_batch_first")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = tensor<string, []>("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> lstm_out_batch_first_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("lstm_out_batch_first_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [1]> hn_axes_0 = const()[name = tensor<string, []>("hn_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> lstm_out_batch_first_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("lstm_out_batch_first_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128]> lstm_out_batch_first_1_to_fp16 = cast(dtype = lstm_out_batch_first_1_to_fp16_dtype_0, x = lstm_out_batch_first_1)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 1, 128]> hn_cast_fp16 = expand_dims(axes = hn_axes_0, x = lstm_out_batch_first_1_to_fp16)[name = tensor<string, []>("hn_cast_fp16")];
+            tensor<int32, [1]> cn_axes_0 = const()[name = tensor<string, []>("cn_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> lstm_out_batch_first_2_to_fp16_dtype_0 = const()[name = tensor<string, []>("lstm_out_batch_first_2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128]> lstm_out_batch_first_2_to_fp16 = cast(dtype = lstm_out_batch_first_2_to_fp16_dtype_0, x = lstm_out_batch_first_2)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1, 128]> cn_cast_fp16 = expand_dims(axes = cn_axes_0, x = lstm_out_batch_first_2_to_fp16)[name = tensor<string, []>("cn_cast_fp16")];
+            tensor<fp16, [1, 1, 128]> lstm_out_batch_first_0_to_fp16 = cast(dtype = lstm_out_batch_first_0_to_fp16_dtype_0, x = lstm_out_batch_first_0)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 128, 1]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = lstm_out_batch_first_0_to_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 128, 1]> input_23_cast_fp16 = relu(x = transpose_1_cast_fp16)[name = tensor<string, []>("input_23_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_final_conv_weight_to_fp16 = const()[name = tensor<string, []>("decoder_final_conv_weight_to_fp16"), val = tensor<fp16, [1, 128, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(881984)))];
+            tensor<fp16, [1]> decoder_final_conv_bias_to_fp16 = const()[name = tensor<string, []>("decoder_final_conv_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.dfp-5])];
+            tensor<fp16, [1, 1, 1]> input_cast_fp16 = conv(bias = decoder_final_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_final_conv_weight_to_fp16, x = input_23_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_124_cast_fp16 = sigmoid(x = input_cast_fp16)[name = tensor<string, []>("op_124_cast_fp16")];
+            tensor<string, []> var_124_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_124_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("op_125_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128]> var_125_cast_fp16 = squeeze(axes = var_125_axes_0, x = hn_cast_fp16)[name = tensor<string, []>("op_125_cast_fp16")];
+            tensor<string, []> var_125_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_125_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_126_axes_0 = const()[name = tensor<string, []>("op_126_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128]> var_126_cast_fp16 = squeeze(axes = var_126_axes_0, x = cn_cast_fp16)[name = tensor<string, []>("op_126_cast_fp16")];
+            tensor<string, []> var_126_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_126_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 128]> new_cell_state = cast(dtype = var_126_cast_fp16_to_fp32_dtype_0, x = var_126_cast_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<fp32, [1, 128]> new_hidden_state = cast(dtype = var_125_cast_fp16_to_fp32_dtype_0, x = var_125_cast_fp16)[name = tensor<string, []>("cast_1")];
+            tensor<fp32, [1, 1, 1]> vad_output = cast(dtype = var_124_cast_fp16_to_fp32_dtype_0, x = var_124_cast_fp16)[name = tensor<string, []>("cast_2")];
+        } -> (vad_output, new_hidden_state, new_cell_state);
+}

silero-vad-unified-v6.0.0.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:853cf34740d3f5061f977ebe2976f7c921b064261c9c4753b3a1196f2dba42b4
+size 882304

silero-vad-unified-v6.0.0.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:532a63e7db357d739761635285c295b6d440a07d081951e914351efc1f6fd8d9
+size 22122

silero-vad-unified-v6.0.0.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:853cf34740d3f5061f977ebe2976f7c921b064261c9c4753b3a1196f2dba42b4
+size 882304

silero-vad-unified-v6.0.0.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "18F41392-7D2B-413D-B76B-A0C5C5514516": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "40D9B5FB-A936-4C45-9B1C-E7E6177C06D0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "18F41392-7D2B-413D-B76B-A0C5C5514516"
+}

silero_vad.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35c6d0bd3f8dd431fed72221005853ffe3621af1b550951093c41d0b918d210e
+size 243

silero_vad.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca7f6a0ab7a349477fed1864e6cf7cb6adf611f017c0c5f0218c694d25e1434a
+size 422

silero_vad.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "shortDescription" : "VAD with SE modules trained on MUSAN (86.47% accuracy)",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "vad_probability",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "2.0",
+    "modelParameters" : [
+
+    ],
+    "author" : "Silero VAD with Trained SE Modules",
+    "specificationVersion" : 6,
+    "storagePrecision" : "Mixed (Float16, Float32)",
+    "mlProgramOperationTypeHistogram" : {
+      "Concat" : 4,
+      "Lstm" : 1,
+      "Linear" : 14,
+      "SliceByIndex" : 3,
+      "LayerNorm" : 1,
+      "Pow" : 6,
+      "Stack" : 1,
+      "Transpose" : 3,
+      "Relu" : 9,
+      "ReduceMean" : 5,
+      "Cast" : 4,
+      "Reshape" : 8,
+      "Add" : 6,
+      "Sqrt" : 3,
+      "Sigmoid" : 5,
+      "Mul" : 5,
+      "Conv" : 5,
+      "Squeeze" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "12.0",
+      "tvOS" : "15.0",
+      "visionOS" : "1.0",
+      "watchOS" : "8.0",
+      "iOS" : "15.0",
+      "macCatalyst" : "15.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "audio_chunk",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "8.3.0"
+    },
+    "generatedClassName" : "silero_vad_se_trained",
+    "method" : "predict"
+  }
+]

silero_vad.mlmodelc/model.mil

@@ -0,0 +1,211 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios15>(tensor<fp32, [1, 512]> audio_chunk) {
+            tensor<int32, [2]> frame_1_begin_0 = const()[name = tensor<string, []>("frame_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> frame_1_end_0 = const()[name = tensor<string, []>("frame_1_end_0"), val = tensor<int32, [2]>([1, 256])];
+            tensor<bool, [2]> frame_1_end_mask_0 = const()[name = tensor<string, []>("frame_1_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, [1, 512]> audio_chunk_to_fp16 = cast(dtype = audio_chunk_to_fp16_dtype_0, x = audio_chunk)[name = tensor<string, []>("cast_11")];
+            tensor<fp16, [1, 256]> frame_1_cast_fp16 = slice_by_index(begin = frame_1_begin_0, end = frame_1_end_0, end_mask = frame_1_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_1_cast_fp16")];
+            tensor<int32, [2]> frame_3_begin_0 = const()[name = tensor<string, []>("frame_3_begin_0"), val = tensor<int32, [2]>([0, 128])];
+            tensor<int32, [2]> frame_3_end_0 = const()[name = tensor<string, []>("frame_3_end_0"), val = tensor<int32, [2]>([1, 384])];
+            tensor<bool, [2]> frame_3_end_mask_0 = const()[name = tensor<string, []>("frame_3_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, 256]> frame_3_cast_fp16 = slice_by_index(begin = frame_3_begin_0, end = frame_3_end_0, end_mask = frame_3_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_3_cast_fp16")];
+            tensor<int32, [2]> frame_begin_0 = const()[name = tensor<string, []>("frame_begin_0"), val = tensor<int32, [2]>([0, 256])];
+            tensor<int32, [2]> frame_end_0 = const()[name = tensor<string, []>("frame_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> frame_end_mask_0 = const()[name = tensor<string, []>("frame_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, 256]> frame_cast_fp16 = slice_by_index(begin = frame_begin_0, end = frame_end_0, end_mask = frame_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_cast_fp16")];
+            tensor<fp16, [129, 256]> var_26_to_fp16 = const()[name = tensor<string, []>("op_26_to_fp16"), val = tensor<fp16, [129, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [129]> var_38_bias_0_to_fp16 = const()[name = tensor<string, []>("op_38_bias_0_to_fp16"), val = tensor<fp16, [129]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66176)))];
+            tensor<fp16, [1, 129]> var_38_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_26_to_fp16, x = frame_1_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<fp16, [129, 256]> var_29_to_fp16 = const()[name = tensor<string, []>("op_29_to_fp16"), val = tensor<fp16, [129, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66560)))];
+            tensor<fp16, [1, 129]> var_40_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_29_to_fp16, x = frame_1_cast_fp16)[name = tensor<string, []>("op_40_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_41_cast_fp16 = pow(x = var_38_cast_fp16, y = var_20_promoted_to_fp16)[name = tensor<string, []>("op_41_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_42_cast_fp16 = pow(x = var_40_cast_fp16, y = var_20_promoted_1_to_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
+            tensor<fp16, [1, 129]> var_43_cast_fp16 = add(x = var_41_cast_fp16, y = var_42_cast_fp16)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<fp16, []> var_44_to_fp16 = const()[name = tensor<string, []>("op_44_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_45_cast_fp16 = add(x = var_43_cast_fp16, y = var_44_to_fp16)[name = tensor<string, []>("op_45_cast_fp16")];
+            tensor<fp16, [1, 129]> var_46_cast_fp16 = sqrt(x = var_45_cast_fp16)[name = tensor<string, []>("op_46_cast_fp16")];
+            tensor<fp16, [1, 129]> var_48_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_26_to_fp16, x = frame_3_cast_fp16)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<fp16, [1, 129]> var_50_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_29_to_fp16, x = frame_3_cast_fp16)[name = tensor<string, []>("op_50_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_51_cast_fp16 = pow(x = var_48_cast_fp16, y = var_20_promoted_2_to_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_3_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_3_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_52_cast_fp16 = pow(x = var_50_cast_fp16, y = var_20_promoted_3_to_fp16)[name = tensor<string, []>("op_52_cast_fp16")];
+            tensor<fp16, [1, 129]> var_53_cast_fp16 = add(x = var_51_cast_fp16, y = var_52_cast_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
+            tensor<fp16, []> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_55_cast_fp16 = add(x = var_53_cast_fp16, y = var_54_to_fp16)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<fp16, [1, 129]> var_56_cast_fp16 = sqrt(x = var_55_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
+            tensor<fp16, [1, 129]> var_58_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_26_to_fp16, x = frame_cast_fp16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<fp16, [1, 129]> var_60_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = var_29_to_fp16, x = frame_cast_fp16)[name = tensor<string, []>("op_60_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_4_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_4_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_61_cast_fp16 = pow(x = var_58_cast_fp16, y = var_20_promoted_4_to_fp16)[name = tensor<string, []>("op_61_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_5_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_5_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_62_cast_fp16 = pow(x = var_60_cast_fp16, y = var_20_promoted_5_to_fp16)[name = tensor<string, []>("op_62_cast_fp16")];
+            tensor<fp16, [1, 129]> var_63_cast_fp16 = add(x = var_61_cast_fp16, y = var_62_cast_fp16)[name = tensor<string, []>("op_63_cast_fp16")];
+            tensor<fp16, []> var_64_to_fp16 = const()[name = tensor<string, []>("op_64_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_65_cast_fp16 = add(x = var_63_cast_fp16, y = var_64_to_fp16)[name = tensor<string, []>("op_65_cast_fp16")];
+            tensor<fp16, [1, 129]> magnitude_cast_fp16 = sqrt(x = var_65_cast_fp16)[name = tensor<string, []>("magnitude_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 = (var_46_cast_fp16, var_56_cast_fp16, magnitude_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]> vad_encoder_encoder_0_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_conv_weight_to_fp16"), val = tensor<fp16, [128, 129, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(132672)))];
+            tensor<fp16, [128]> vad_encoder_encoder_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231808)))];
+            tensor<fp16, [1, 128, 3]> x_1_cast_fp16 = conv(bias = vad_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 = vad_encoder_encoder_0_conv_weight_to_fp16, 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_92 = const()[name = tensor<string, []>("op_92"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_3_cast_fp16 = reshape(shape = var_92, x = concat_0_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [16, 128]> vad_encoder_encoder_0_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(232128)))];
+            tensor<fp16, [16]> vad_encoder_encoder_0_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236288)))];
+            tensor<fp16, [1, 16]> linear_0_cast_fp16 = linear(bias = vad_encoder_encoder_0_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_0_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(236416)))];
+            tensor<fp16, [128]> vad_encoder_encoder_0_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_se_fc2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240576)))];
+            tensor<fp16, [1, 128]> linear_1_cast_fp16 = linear(bias = vad_encoder_encoder_0_se_fc2_bias_to_fp16, weight = vad_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_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_3_cast_fp16 = reshape(shape = var_102, 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]> vad_encoder_encoder_1_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_conv_weight_to_fp16"), val = tensor<fp16, [64, 128, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240896)))];
+            tensor<fp16, [64]> vad_encoder_encoder_1_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290112)))];
+            tensor<fp16, [1, 64, 3]> x_3_cast_fp16 = conv(bias = vad_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 = vad_encoder_encoder_1_conv_weight_to_fp16, 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_121 = const()[name = tensor<string, []>("op_121"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_15_cast_fp16 = reshape(shape = var_121, x = concat_1_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, [8, 64]> vad_encoder_encoder_1_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(290304)))];
+            tensor<fp16, [8]> vad_encoder_encoder_1_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x1.00cp-9, 0x1.6dcp-5, 0x1.82cp-5, 0x1.054p-5, 0x1.8a4p-4, 0x1.f88p-7, 0x1.234p-5, 0x1.514p-5])];
+            tensor<fp16, [1, 8]> linear_2_cast_fp16 = linear(bias = vad_encoder_encoder_1_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_1_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(291392)))];
+            tensor<fp16, [64]> vad_encoder_encoder_1_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292480)))];
+            tensor<fp16, [1, 64]> linear_3_cast_fp16 = linear(bias = vad_encoder_encoder_1_se_fc2_bias_to_fp16, weight = vad_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_131 = const()[name = tensor<string, []>("op_131"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_7_cast_fp16 = reshape(shape = var_131, 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]> vad_encoder_encoder_2_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_conv_weight_to_fp16"), val = tensor<fp16, [64, 64, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292672)))];
+            tensor<fp16, [64]> vad_encoder_encoder_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317312)))];
+            tensor<fp16, [1, 64, 3]> x_5_cast_fp16 = conv(bias = vad_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 = vad_encoder_encoder_2_conv_weight_to_fp16, 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_150 = const()[name = tensor<string, []>("op_150"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_27_cast_fp16 = reshape(shape = var_150, x = concat_2_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [8, 64]> vad_encoder_encoder_2_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(317504)))];
+            tensor<fp16, [8]> vad_encoder_encoder_2_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x1.41cp-7, 0x1.4e8p-5, 0x1.7ccp-5, 0x1.554p-5, 0x1.d8cp-5, -0x1.34p-11, 0x1.2f8p-5, 0x1.0ap-5])];
+            tensor<fp16, [1, 8]> linear_4_cast_fp16 = linear(bias = vad_encoder_encoder_2_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_2_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(318592)))];
+            tensor<fp16, [64]> vad_encoder_encoder_2_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319680)))];
+            tensor<fp16, [1, 64]> linear_5_cast_fp16 = linear(bias = vad_encoder_encoder_2_se_fc2_bias_to_fp16, weight = vad_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_160 = const()[name = tensor<string, []>("op_160"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_11_cast_fp16 = reshape(shape = var_160, 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]> vad_encoder_encoder_3_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_conv_weight_to_fp16"), val = tensor<fp16, [128, 64, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319872)))];
+            tensor<fp16, [128]> vad_encoder_encoder_3_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369088)))];
+            tensor<fp16, [1, 128, 3]> x_7_cast_fp16 = conv(bias = vad_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 = vad_encoder_encoder_3_conv_weight_to_fp16, 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_179 = const()[name = tensor<string, []>("op_179"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_39_cast_fp16 = reshape(shape = var_179, x = concat_3_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [16, 128]> vad_encoder_encoder_3_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(369408)))];
+            tensor<fp16, [16]> vad_encoder_encoder_3_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373568)))];
+            tensor<fp16, [1, 16]> linear_6_cast_fp16 = linear(bias = vad_encoder_encoder_3_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_3_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(373696)))];
+            tensor<fp16, [128]> vad_encoder_encoder_3_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_se_fc2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377856)))];
+            tensor<fp16, [1, 128]> linear_7_cast_fp16 = linear(bias = vad_encoder_encoder_3_se_fc2_bias_to_fp16, weight = vad_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_189 = const()[name = tensor<string, []>("op_189"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_cast_fp16 = reshape(shape = var_189, 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_6_perm_0 = const()[name = tensor<string, []>("transpose_6_perm_0"), val = tensor<int32, [3]>([2, 0, 1])];
+            tensor<string, []> transpose_6_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_6_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [512]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378176)))];
+            tensor<fp32, [512, 128]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380288)))];
+            tensor<fp32, [512, 128]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(642496)))];
+            tensor<fp32, [1, 128]> input_49_batch_first_lstm_h0_squeeze = const()[name = tensor<string, []>("input_49_batch_first_lstm_h0_squeeze"), val = tensor<fp32, [1, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(904704)))];
+            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, [1, 1, 128]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_9")];
+            tensor<fp32, [1, 1, 128]> transpose_6_cast_fp16_to_fp32 = cast(dtype = transpose_6_cast_fp16_to_fp32_dtype_0, x = transpose_6_cast_fp16)[name = tensor<string, []>("cast_10")];
+            tensor<fp32, [1, 1, 128]> input_49_batch_first_0, tensor<fp32, [1, 128]> input_49_batch_first_1, tensor<fp32, [1, 128]> input_49_batch_first_2 = lstm(activation = input_49_batch_first_activation_0, bias = concat_4, cell_activation = input_49_batch_first_cell_activation_0, direction = input_49_batch_first_direction_0, initial_c = input_49_batch_first_lstm_h0_squeeze, initial_h = input_49_batch_first_lstm_h0_squeeze, output_sequence = input_49_batch_first_output_sequence_0, recurrent_activation = input_49_batch_first_recurrent_activation_0, weight_hh = concat_6, weight_ih = concat_5, x = transpose_6_cast_fp16_to_fp32)[name = tensor<string, []>("input_49_batch_first")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<string, []> input_49_batch_first_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_49_batch_first_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [1]> var_216_axes_0 = const()[name = tensor<string, []>("op_216_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> vad_decoder_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("vad_decoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(905280)))];
+            tensor<fp16, [128]> vad_decoder_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("vad_decoder_layer_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(905600)))];
+            tensor<fp16, []> var_5_to_fp16 = const()[name = tensor<string, []>("op_5_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 128]> input_49_batch_first_0_to_fp16 = cast(dtype = input_49_batch_first_0_to_fp16_dtype_0, x = input_49_batch_first_0)[name = tensor<string, []>("cast_9")];
+            tensor<fp16, [1, 1, 128]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = input_49_batch_first_0_to_fp16)[name = tensor<string, []>("transpose_8")];
+            tensor<fp16, [1, 1, 128]> var_216_cast_fp16 = layer_norm(axes = var_216_axes_0, beta = vad_decoder_layer_norm_bias_to_fp16, epsilon = var_5_to_fp16, gamma = vad_decoder_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
+            tensor<fp16, []> var_217_to_fp16 = const()[name = tensor<string, []>("op_217_to_fp16"), val = tensor<fp16, []>(0x1.334p-3)];
+            tensor<fp16, [1, 1, 128]> x_cast_fp16 = mul(x = var_216_cast_fp16, y = var_217_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_7")];
+            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]> vad_decoder_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_decoder_conv_weight_to_fp16"), val = tensor<fp16, [1, 128, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(905920)))];
+            tensor<fp16, [1]> vad_decoder_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_decoder_conv_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.dfp-5])];
+            tensor<fp16, [1, 1, 1]> input_cast_fp16 = conv(bias = vad_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 = vad_decoder_conv_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_229_cast_fp16 = sigmoid(x = input_cast_fp16)[name = tensor<string, []>("op_229_cast_fp16")];
+            tensor<int32, [1]> var_230_axes_0 = const()[name = tensor<string, []>("op_230_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_230_cast_fp16 = squeeze(axes = var_230_axes_0, x = var_229_cast_fp16)[name = tensor<string, []>("op_230_cast_fp16")];
+            tensor<string, []> var_230_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_230_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1]> vad_probability = cast(dtype = var_230_cast_fp16_to_fp32_dtype_0, x = var_230_cast_fp16)[name = tensor<string, []>("cast_8")];
+        } -> (vad_probability);
+}

silero_vad.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45846d0738d3bf5e4b6e9e7d2fddda7b1ad07da33d473f0405e51d3b6c4c11a9
+size 906240

silero_vad_se_trained.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74b401d76333292a0cafa1024f98f4e4e838427e67f215b65c95ddeed6a2d16e
+size 32557

silero_vad_se_trained.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45846d0738d3bf5e4b6e9e7d2fddda7b1ad07da33d473f0405e51d3b6c4c11a9
+size 906240

silero_vad_se_trained.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "5162830C-61A9-43CB-85D2-5E27DE992B03": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "C9FCCBCB-EEAA-416F-B703-903A78B678E1": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "C9FCCBCB-EEAA-416F-B703-903A78B678E1"
+}

silero_vad_se_trained_4bit.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed0b15a6463ed93cf4e8ee8ad21d5df4e6c54e1db074b646bcd88b9acfa517a4
+size 243

silero_vad_se_trained_4bit.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d430fe5a090b2ae6424d79b458c3ecc7899d65b84df6ad0455df0bb0b378822
+size 422

silero_vad_se_trained_4bit.mlmodelc/metadata.json

@@ -0,0 +1,81 @@
+[
+  {
+    "shortDescription" : "VAD with SE modules trained on MUSAN (86.47% accuracy)",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "vad_probability",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "2.0",
+    "modelParameters" : [
+
+    ],
+    "author" : "Silero VAD with Trained SE Modules",
+    "specificationVersion" : 7,
+    "storagePrecision" : "Mixed (Float16, Float32, Palettized (4 bits))",
+    "mlProgramOperationTypeHistogram" : {
+      "Concat" : 4,
+      "Ios16.cast" : 4,
+      "Ios16.mul" : 5,
+      "Ios16.layerNorm" : 1,
+      "SliceByIndex" : 3,
+      "Ios16.constexprLutToDense" : 8,
+      "Ios16.sqrt" : 3,
+      "Stack" : 1,
+      "Transpose" : 3,
+      "Ios16.sigmoid" : 5,
+      "Ios16.conv" : 5,
+      "Ios16.add" : 6,
+      "Ios16.linear" : 14,
+      "Ios16.reduceMean" : 5,
+      "Ios16.relu" : 9,
+      "Ios16.lstm" : 1,
+      "Ios16.pow" : 6,
+      "Squeeze" : 1,
+      "Ios16.reshape" : 8
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "13.0",
+      "tvOS" : "16.0",
+      "visionOS" : "1.0",
+      "watchOS" : "9.0",
+      "iOS" : "16.0",
+      "macCatalyst" : "16.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "audio_chunk",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.0"
+    },
+    "generatedClassName" : "silero_vad_se_trained_palettize_4bit",
+    "method" : "predict"
+  }
+]

silero_vad_se_trained_4bit.mlmodelc/model.mil

@@ -0,0 +1,211 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios16>(tensor<fp32, [1, 512]> audio_chunk) {
+            tensor<int32, [2]> frame_1_begin_0 = const()[name = tensor<string, []>("frame_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> frame_1_end_0 = const()[name = tensor<string, []>("frame_1_end_0"), val = tensor<int32, [2]>([1, 256])];
+            tensor<bool, [2]> frame_1_end_mask_0 = const()[name = tensor<string, []>("frame_1_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, [1, 512]> audio_chunk_to_fp16 = cast(dtype = audio_chunk_to_fp16_dtype_0, x = audio_chunk)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 256]> frame_1_cast_fp16 = slice_by_index(begin = frame_1_begin_0, end = frame_1_end_0, end_mask = frame_1_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_1_cast_fp16")];
+            tensor<int32, [2]> frame_3_begin_0 = const()[name = tensor<string, []>("frame_3_begin_0"), val = tensor<int32, [2]>([0, 128])];
+            tensor<int32, [2]> frame_3_end_0 = const()[name = tensor<string, []>("frame_3_end_0"), val = tensor<int32, [2]>([1, 384])];
+            tensor<bool, [2]> frame_3_end_mask_0 = const()[name = tensor<string, []>("frame_3_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, 256]> frame_3_cast_fp16 = slice_by_index(begin = frame_3_begin_0, end = frame_3_end_0, end_mask = frame_3_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_3_cast_fp16")];
+            tensor<int32, [2]> frame_begin_0 = const()[name = tensor<string, []>("frame_begin_0"), val = tensor<int32, [2]>([0, 256])];
+            tensor<int32, [2]> frame_end_0 = const()[name = tensor<string, []>("frame_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> frame_end_mask_0 = const()[name = tensor<string, []>("frame_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, 256]> frame_cast_fp16 = slice_by_index(begin = frame_begin_0, end = frame_end_0, end_mask = frame_end_mask_0, x = audio_chunk_to_fp16)[name = tensor<string, []>("frame_cast_fp16")];
+            tensor<fp16, [129, 256]> op_26_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16512]>(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, []>(16640))), name = tensor<string, []>("op_26_to_fp16_palettized"), shape = tensor<uint32, [2]>([129, 256])];
+            tensor<fp16, [129]> var_38_bias_0_to_fp16 = const()[name = tensor<string, []>("op_38_bias_0_to_fp16"), val = tensor<fp16, [129]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16768)))];
+            tensor<fp16, [1, 129]> var_38_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_26_to_fp16_palettized, x = frame_1_cast_fp16)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<fp16, [129, 256]> op_29_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17152))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33728))), name = tensor<string, []>("op_29_to_fp16_palettized"), shape = tensor<uint32, [2]>([129, 256])];
+            tensor<fp16, [1, 129]> var_40_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_29_to_fp16_palettized, x = frame_1_cast_fp16)[name = tensor<string, []>("op_40_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_41_cast_fp16 = pow(x = var_38_cast_fp16, y = var_20_promoted_to_fp16)[name = tensor<string, []>("op_41_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_1_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_1_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_42_cast_fp16 = pow(x = var_40_cast_fp16, y = var_20_promoted_1_to_fp16)[name = tensor<string, []>("op_42_cast_fp16")];
+            tensor<fp16, [1, 129]> var_43_cast_fp16 = add(x = var_41_cast_fp16, y = var_42_cast_fp16)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<fp16, []> var_44_to_fp16 = const()[name = tensor<string, []>("op_44_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_45_cast_fp16 = add(x = var_43_cast_fp16, y = var_44_to_fp16)[name = tensor<string, []>("op_45_cast_fp16")];
+            tensor<fp16, [1, 129]> var_46_cast_fp16 = sqrt(x = var_45_cast_fp16)[name = tensor<string, []>("op_46_cast_fp16")];
+            tensor<fp16, [1, 129]> var_48_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_26_to_fp16_palettized, x = frame_3_cast_fp16)[name = tensor<string, []>("op_48_cast_fp16")];
+            tensor<fp16, [1, 129]> var_50_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_29_to_fp16_palettized, x = frame_3_cast_fp16)[name = tensor<string, []>("op_50_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_2_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_2_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_51_cast_fp16 = pow(x = var_48_cast_fp16, y = var_20_promoted_2_to_fp16)[name = tensor<string, []>("op_51_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_3_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_3_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_52_cast_fp16 = pow(x = var_50_cast_fp16, y = var_20_promoted_3_to_fp16)[name = tensor<string, []>("op_52_cast_fp16")];
+            tensor<fp16, [1, 129]> var_53_cast_fp16 = add(x = var_51_cast_fp16, y = var_52_cast_fp16)[name = tensor<string, []>("op_53_cast_fp16")];
+            tensor<fp16, []> var_54_to_fp16 = const()[name = tensor<string, []>("op_54_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_55_cast_fp16 = add(x = var_53_cast_fp16, y = var_54_to_fp16)[name = tensor<string, []>("op_55_cast_fp16")];
+            tensor<fp16, [1, 129]> var_56_cast_fp16 = sqrt(x = var_55_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
+            tensor<fp16, [1, 129]> var_58_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_26_to_fp16_palettized, x = frame_cast_fp16)[name = tensor<string, []>("op_58_cast_fp16")];
+            tensor<fp16, [1, 129]> var_60_cast_fp16 = linear(bias = var_38_bias_0_to_fp16, weight = op_29_to_fp16_palettized, x = frame_cast_fp16)[name = tensor<string, []>("op_60_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_4_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_4_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_61_cast_fp16 = pow(x = var_58_cast_fp16, y = var_20_promoted_4_to_fp16)[name = tensor<string, []>("op_61_cast_fp16")];
+            tensor<fp16, []> var_20_promoted_5_to_fp16 = const()[name = tensor<string, []>("op_20_promoted_5_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
+            tensor<fp16, [1, 129]> var_62_cast_fp16 = pow(x = var_60_cast_fp16, y = var_20_promoted_5_to_fp16)[name = tensor<string, []>("op_62_cast_fp16")];
+            tensor<fp16, [1, 129]> var_63_cast_fp16 = add(x = var_61_cast_fp16, y = var_62_cast_fp16)[name = tensor<string, []>("op_63_cast_fp16")];
+            tensor<fp16, []> var_64_to_fp16 = const()[name = tensor<string, []>("op_64_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
+            tensor<fp16, [1, 129]> var_65_cast_fp16 = add(x = var_63_cast_fp16, y = var_64_to_fp16)[name = tensor<string, []>("op_65_cast_fp16")];
+            tensor<fp16, [1, 129]> magnitude_cast_fp16 = sqrt(x = var_65_cast_fp16)[name = tensor<string, []>("magnitude_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 = (var_46_cast_fp16, var_56_cast_fp16, magnitude_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]> vad_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, []>(33856))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58688))), name = tensor<string, []>("vad_encoder_encoder_0_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 129, 3])];
+            tensor<fp16, [128]> vad_encoder_encoder_0_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58816)))];
+            tensor<fp16, [1, 128, 3]> x_1_cast_fp16 = conv(bias = vad_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 = vad_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_92 = const()[name = tensor<string, []>("op_92"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_3_cast_fp16 = reshape(shape = var_92, x = concat_0_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [16, 128]> vad_encoder_encoder_0_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(59136)))];
+            tensor<fp16, [16]> vad_encoder_encoder_0_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63296)))];
+            tensor<fp16, [1, 16]> linear_0_cast_fp16 = linear(bias = vad_encoder_encoder_0_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_0_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(63424)))];
+            tensor<fp16, [128]> vad_encoder_encoder_0_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_0_se_fc2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67584)))];
+            tensor<fp16, [1, 128]> linear_1_cast_fp16 = linear(bias = vad_encoder_encoder_0_se_fc2_bias_to_fp16, weight = vad_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_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_3_cast_fp16 = reshape(shape = var_102, 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]> vad_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, []>(67904))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80256))), name = tensor<string, []>("vad_encoder_encoder_1_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([64, 128, 3])];
+            tensor<fp16, [64]> vad_encoder_encoder_1_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80384)))];
+            tensor<fp16, [1, 64, 3]> x_3_cast_fp16 = conv(bias = vad_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 = vad_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_121 = const()[name = tensor<string, []>("op_121"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_15_cast_fp16 = reshape(shape = var_121, x = concat_1_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<fp16, [8, 64]> vad_encoder_encoder_1_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(80576)))];
+            tensor<fp16, [8]> vad_encoder_encoder_1_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x1.00cp-9, 0x1.6dcp-5, 0x1.82cp-5, 0x1.054p-5, 0x1.8a4p-4, 0x1.f88p-7, 0x1.234p-5, 0x1.514p-5])];
+            tensor<fp16, [1, 8]> linear_2_cast_fp16 = linear(bias = vad_encoder_encoder_1_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_1_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(81664)))];
+            tensor<fp16, [64]> vad_encoder_encoder_1_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_1_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82752)))];
+            tensor<fp16, [1, 64]> linear_3_cast_fp16 = linear(bias = vad_encoder_encoder_1_se_fc2_bias_to_fp16, weight = vad_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_131 = const()[name = tensor<string, []>("op_131"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_7_cast_fp16 = reshape(shape = var_131, 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]> vad_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, []>(82944))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89152))), name = tensor<string, []>("vad_encoder_encoder_2_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([64, 64, 3])];
+            tensor<fp16, [64]> vad_encoder_encoder_2_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_conv_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89280)))];
+            tensor<fp16, [1, 64, 3]> x_5_cast_fp16 = conv(bias = vad_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 = vad_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_150 = const()[name = tensor<string, []>("op_150"), val = tensor<int32, [2]>([1, 64])];
+            tensor<fp16, [1, 64]> input_27_cast_fp16 = reshape(shape = var_150, x = concat_2_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [8, 64]> vad_encoder_encoder_2_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(89472)))];
+            tensor<fp16, [8]> vad_encoder_encoder_2_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_se_fc1_bias_to_fp16"), val = tensor<fp16, [8]>([0x1.41cp-7, 0x1.4e8p-5, 0x1.7ccp-5, 0x1.554p-5, 0x1.d8cp-5, -0x1.34p-11, 0x1.2f8p-5, 0x1.0ap-5])];
+            tensor<fp16, [1, 8]> linear_4_cast_fp16 = linear(bias = vad_encoder_encoder_2_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_2_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(90560)))];
+            tensor<fp16, [64]> vad_encoder_encoder_2_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_2_se_fc2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(91648)))];
+            tensor<fp16, [1, 64]> linear_5_cast_fp16 = linear(bias = vad_encoder_encoder_2_se_fc2_bias_to_fp16, weight = vad_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_160 = const()[name = tensor<string, []>("op_160"), val = tensor<int32, [3]>([1, 64, 1])];
+            tensor<fp16, [1, 64, 1]> y_11_cast_fp16 = reshape(shape = var_160, 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]> vad_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, []>(91840))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104192))), name = tensor<string, []>("vad_encoder_encoder_3_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 64, 3])];
+            tensor<fp16, [128]> vad_encoder_encoder_3_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(104320)))];
+            tensor<fp16, [1, 128, 3]> x_7_cast_fp16 = conv(bias = vad_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 = vad_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_179 = const()[name = tensor<string, []>("op_179"), val = tensor<int32, [2]>([1, 128])];
+            tensor<fp16, [1, 128]> input_39_cast_fp16 = reshape(shape = var_179, x = concat_3_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [16, 128]> vad_encoder_encoder_3_se_fc1_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(104640)))];
+            tensor<fp16, [16]> vad_encoder_encoder_3_se_fc1_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_se_fc1_bias_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108800)))];
+            tensor<fp16, [1, 16]> linear_6_cast_fp16 = linear(bias = vad_encoder_encoder_3_se_fc1_bias_to_fp16, weight = vad_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]> vad_encoder_encoder_3_se_fc2_weight_to_fp16 = const()[name = tensor<string, []>("vad_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, []>(108928)))];
+            tensor<fp16, [128]> vad_encoder_encoder_3_se_fc2_bias_to_fp16 = const()[name = tensor<string, []>("vad_encoder_encoder_3_se_fc2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113088)))];
+            tensor<fp16, [1, 128]> linear_7_cast_fp16 = linear(bias = vad_encoder_encoder_3_se_fc2_bias_to_fp16, weight = vad_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_189 = const()[name = tensor<string, []>("op_189"), val = tensor<int32, [3]>([1, 128, 1])];
+            tensor<fp16, [1, 128, 1]> y_cast_fp16 = reshape(shape = var_189, 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_6_perm_0 = const()[name = tensor<string, []>("transpose_6_perm_0"), val = tensor<int32, [3]>([2, 0, 1])];
+            tensor<string, []> transpose_6_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_6_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [512]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113408)))];
+            tensor<fp32, [512, 128]> concat_5_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115520))), lut = tensor<fp32, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148352))), name = tensor<string, []>("concat_5_palettized"), shape = tensor<uint32, [2]>([512, 128])];
+            tensor<fp32, [512, 128]> concat_6_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148480))), lut = tensor<fp32, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181312))), name = tensor<string, []>("concat_6_palettized"), shape = tensor<uint32, [2]>([512, 128])];
+            tensor<fp32, [1, 128]> input_49_batch_first_lstm_h0_squeeze = const()[name = tensor<string, []>("input_49_batch_first_lstm_h0_squeeze"), val = tensor<fp32, [1, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(181440)))];
+            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, [1, 1, 128]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 1, 128]> transpose_6_cast_fp16_to_fp32 = cast(dtype = transpose_6_cast_fp16_to_fp32_dtype_0, x = transpose_6_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [1, 1, 128]> input_49_batch_first_0, tensor<fp32, [1, 128]> input_49_batch_first_1, tensor<fp32, [1, 128]> input_49_batch_first_2 = lstm(activation = input_49_batch_first_activation_0, bias = concat_4, cell_activation = input_49_batch_first_cell_activation_0, direction = input_49_batch_first_direction_0, initial_c = input_49_batch_first_lstm_h0_squeeze, initial_h = input_49_batch_first_lstm_h0_squeeze, output_sequence = input_49_batch_first_output_sequence_0, recurrent_activation = input_49_batch_first_recurrent_activation_0, weight_hh = concat_6_palettized, weight_ih = concat_5_palettized, x = transpose_6_cast_fp16_to_fp32)[name = tensor<string, []>("input_49_batch_first")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<string, []> input_49_batch_first_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("input_49_batch_first_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [1]> var_216_axes_0 = const()[name = tensor<string, []>("op_216_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [128]> vad_decoder_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("vad_decoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182016)))];
+            tensor<fp16, [128]> vad_decoder_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("vad_decoder_layer_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182336)))];
+            tensor<fp16, []> var_5_to_fp16 = const()[name = tensor<string, []>("op_5_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [1, 1, 128]> input_49_batch_first_0_to_fp16 = cast(dtype = input_49_batch_first_0_to_fp16_dtype_0, x = input_49_batch_first_0)[name = tensor<string, []>("cast_1")];
+            tensor<fp16, [1, 1, 128]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = input_49_batch_first_0_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 128]> var_216_cast_fp16 = layer_norm(axes = var_216_axes_0, beta = vad_decoder_layer_norm_bias_to_fp16, epsilon = var_5_to_fp16, gamma = vad_decoder_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("op_216_cast_fp16")];
+            tensor<fp16, []> var_217_to_fp16 = const()[name = tensor<string, []>("op_217_to_fp16"), val = tensor<fp16, []>(0x1.334p-3)];
+            tensor<fp16, [1, 1, 128]> x_cast_fp16 = mul(x = var_216_cast_fp16, y = var_217_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]> vad_decoder_conv_weight_to_fp16 = const()[name = tensor<string, []>("vad_decoder_conv_weight_to_fp16"), val = tensor<fp16, [1, 128, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182656)))];
+            tensor<fp16, [1]> vad_decoder_conv_bias_to_fp16 = const()[name = tensor<string, []>("vad_decoder_conv_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.dfp-5])];
+            tensor<fp16, [1, 1, 1]> input_cast_fp16 = conv(bias = vad_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 = vad_decoder_conv_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<fp16, [1, 1, 1]> var_229_cast_fp16 = sigmoid(x = input_cast_fp16)[name = tensor<string, []>("op_229_cast_fp16")];
+            tensor<int32, [1]> var_230_axes_0 = const()[name = tensor<string, []>("op_230_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_230_cast_fp16 = squeeze(axes = var_230_axes_0, x = var_229_cast_fp16)[name = tensor<string, []>("op_230_cast_fp16")];
+            tensor<string, []> var_230_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_230_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1]> vad_probability = cast(dtype = var_230_cast_fp16_to_fp32_dtype_0, x = var_230_cast_fp16)[name = tensor<string, []>("cast_0")];
+        } -> (vad_probability);
+}

silero_vad_se_trained_4bit.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f3db1305052551cd81139f0a0de3d5555fa42de357164888f470e2547bb120c
+size 182976