Adding stateful 4bit model as well

silero_vad_stateful_4bit.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a6310ac1abb0ec7a06994d0c46e66c34afdf0fa23d88044eeb695e77011fcb0
+size 243

silero_vad_stateful_4bit.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d945f4a9056d9eeebd6abc43d4d977f7a2da55d7fc6bb5993636ebc728fe371b
+size 529

silero_vad_stateful_4bit.mlmodelc/metadata.json

@@ -0,0 +1,111 @@
+[
+  {
+    "shortDescription" : "Stateful LSTM VAD with proper state management for iOS 17+",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "vad_probability",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 128]",
+        "name" : "h_state_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 128]",
+        "name" : "c_state_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "1.0",
+    "modelParameters" : [
+
+    ],
+    "author" : "Clean Stateful VAD Model",
+    "specificationVersion" : 8,
+    "storagePrecision" : "Mixed (Float16, Palettized (4 bits))",
+    "mlProgramOperationTypeHistogram" : {
+      "Ios16.sigmoid" : 1,
+      "Ios17.squeeze" : 4,
+      "Ios17.cast" : 6,
+      "Ios17.linear" : 4,
+      "Ios17.transpose" : 2,
+      "Ios16.relu" : 2,
+      "Ios16.constexprLutToDense" : 5,
+      "Ios17.lstm" : 1,
+      "Ios17.expandDims" : 3
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "audio_chunk",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 128]",
+        "name" : "h_state_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 128]",
+        "name" : "c_state_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "silero_vad_stateful_4bit_clean",
+    "method" : "predict"
+  }
+]

silero_vad_stateful_4bit.mlmodelc/model.mil

@@ -0,0 +1,60 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 512]> audio_chunk, tensor<fp32, [1, 1, 128]> c_state_in, tensor<fp32, [1, 1, 128]> h_state_in) {
+            tensor<string, []> audio_chunk_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_chunk_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [128, 512]> encoder_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32896))), name = tensor<string, []>("encoder_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 512])];
+            tensor<fp16, [128]> encoder_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_0_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33024)))];
+            tensor<fp16, [1, 512]> audio_chunk_to_fp16 = cast(dtype = audio_chunk_to_fp16_dtype_0, x = audio_chunk)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 128]> linear_0_cast_fp16 = linear(bias = encoder_0_bias_to_fp16, weight = encoder_0_weight_to_fp16_palettized, x = audio_chunk_to_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [1, 128]> input_3_cast_fp16 = relu(x = linear_0_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [64, 128]> encoder_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33344))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37504))), name = tensor<string, []>("encoder_2_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([64, 128])];
+            tensor<fp16, [64]> encoder_2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_2_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37632)))];
+            tensor<fp16, [1, 64]> linear_1_cast_fp16 = linear(bias = encoder_2_bias_to_fp16, weight = encoder_2_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 64]> input_5_cast_fp16 = expand_dims(axes = input_5_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<string, []> h_state_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_state_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> c_state_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_state_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_5_batch_first_transpose_perm_0 = const()[name = tensor<string, []>("input_5_batch_first_transpose_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> lstm_out_batch_first_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> h_state_in_to_fp16 = cast(dtype = h_state_in_to_fp16_dtype_0, x = h_state_in)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 128]> lstm_out_batch_first_lstm_h0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_h0_squeeze_axes_0, x = h_state_in_to_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> lstm_out_batch_first_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> c_state_in_to_fp16 = cast(dtype = c_state_in_to_fp16_dtype_0, x = c_state_in)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 128]> lstm_out_batch_first_lstm_c0_squeeze_cast_fp16 = squeeze(axes = lstm_out_batch_first_lstm_c0_squeeze_axes_0, x = c_state_in_to_fp16)[name = tensor<string, []>("lstm_out_batch_first_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> lstm_out_batch_first_direction_0 = const()[name = tensor<string, []>("lstm_out_batch_first_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> lstm_out_batch_first_output_sequence_0 = const()[name = tensor<string, []>("lstm_out_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> lstm_out_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> lstm_out_batch_first_cell_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> lstm_out_batch_first_activation_0 = const()[name = tensor<string, []>("lstm_out_batch_first_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [512, 64]> concat_1_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37824))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54272))), name = tensor<string, []>("concat_1_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 64])];
+            tensor<fp16, [512, 128]> concat_2_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54400))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87232))), name = tensor<string, []>("concat_2_to_fp16_palettized"), shape = tensor<uint32, [2]>([512, 128])];
+            tensor<fp16, [512]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87360)))];
+            tensor<fp16, [1, 1, 64]> input_5_batch_first_transpose_cast_fp16 = transpose(perm = input_5_batch_first_transpose_perm_0, x = input_5_cast_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 128]> lstm_out_batch_first_cast_fp16_0, tensor<fp16, [1, 128]> lstm_out_batch_first_cast_fp16_1, tensor<fp16, [1, 128]> lstm_out_batch_first_cast_fp16_2 = lstm(activation = lstm_out_batch_first_activation_0, bias = concat_0_to_fp16, cell_activation = lstm_out_batch_first_cell_activation_0, direction = lstm_out_batch_first_direction_0, initial_c = lstm_out_batch_first_lstm_c0_squeeze_cast_fp16, initial_h = lstm_out_batch_first_lstm_h0_squeeze_cast_fp16, output_sequence = lstm_out_batch_first_output_sequence_0, recurrent_activation = lstm_out_batch_first_recurrent_activation_0, weight_hh = concat_2_to_fp16_palettized, weight_ih = concat_1_to_fp16_palettized, x = input_5_batch_first_transpose_cast_fp16)[name = tensor<string, []>("lstm_out_batch_first_cast_fp16")];
+            tensor<int32, [3]> lstm_out_perm_0 = const()[name = tensor<string, []>("lstm_out_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> var_37_axes_0 = const()[name = tensor<string, []>("op_37_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> var_37_cast_fp16 = expand_dims(axes = var_37_axes_0, x = lstm_out_batch_first_cast_fp16_1)[name = tensor<string, []>("op_37_cast_fp16")];
+            tensor<string, []> var_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_37_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_38_axes_0 = const()[name = tensor<string, []>("op_38_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 128]> var_38_cast_fp16 = expand_dims(axes = var_38_axes_0, x = lstm_out_batch_first_cast_fp16_2)[name = tensor<string, []>("op_38_cast_fp16")];
+            tensor<string, []> var_38_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_38_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> lstm_out_cast_fp16 = transpose(perm = lstm_out_perm_0, x = lstm_out_batch_first_cast_fp16_0)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 128]> input_7_cast_fp16 = squeeze(axes = input_7_axes_0, x = lstm_out_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [64, 128]> decoder_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88448))), lut = tensor<fp16, [16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92608))), name = tensor<string, []>("decoder_0_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([64, 128])];
+            tensor<fp16, [64]> decoder_0_bias_to_fp16 = const()[name = tensor<string, []>("decoder_0_bias_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92736)))];
+            tensor<fp16, [1, 64]> linear_2_cast_fp16 = linear(bias = decoder_0_bias_to_fp16, weight = decoder_0_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<fp16, [1, 64]> input_11_cast_fp16 = relu(x = linear_2_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [1, 64]> decoder_2_weight_to_fp16 = const()[name = tensor<string, []>("decoder_2_weight_to_fp16"), val = tensor<fp16, [1, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92928)))];
+            tensor<fp16, [1]> decoder_2_bias_to_fp16 = const()[name = tensor<string, []>("decoder_2_bias_to_fp16"), val = tensor<fp16, [1]>([-0x1.92p-4])];
+            tensor<fp16, [1, 1]> linear_3_cast_fp16 = linear(bias = decoder_2_bias_to_fp16, weight = decoder_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 1]> prob_cast_fp16 = sigmoid(x = linear_3_cast_fp16)[name = tensor<string, []>("prob_cast_fp16")];
+            tensor<int32, [1]> var_56_axes_0 = const()[name = tensor<string, []>("op_56_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1]> var_56_cast_fp16 = squeeze(axes = var_56_axes_0, x = prob_cast_fp16)[name = tensor<string, []>("op_56_cast_fp16")];
+            tensor<string, []> var_56_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_56_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 128]> h_state_out = cast(dtype = var_37_cast_fp16_to_fp32_dtype_0, x = var_37_cast_fp16)[name = tensor<string, []>("cast_0")];
+            tensor<fp32, [1, 1, 128]> c_state_out = cast(dtype = var_38_cast_fp16_to_fp32_dtype_0, x = var_38_cast_fp16)[name = tensor<string, []>("cast_1")];
+            tensor<fp32, [1]> vad_probability = cast(dtype = var_56_cast_fp16_to_fp32_dtype_0, x = var_56_cast_fp16)[name = tensor<string, []>("cast_2")];
+        } -> (vad_probability, h_state_out, c_state_out);
+}

silero_vad_stateful_4bit.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f4767f28d716a87ffa0e446ae9dac70e1ca3708548e65691ae407071b56df8b
+size 93120