Duplicate from FluidInference/parakeet-tdt-ctc-110m-coreml

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

.gitattributes

@@ -0,0 +1,35 @@
+*.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

Decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7174d25586869ffb33e6c519cfcc882bf04b2f0e527b8b84a38fe4d469f618
+size 243

Decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:407846e00c895ee447ee5795eb61c0437b4d5ed8ae33e603173fb3f79d871151
+size 562

Decoder.mlmodelc/metadata.json

@@ -0,0 +1,122 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet TDT-CTC decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 8,
+      "Ios17.lstm" : 1,
+      "Ios17.transpose" : 2,
+      "Ios17.add" : 1,
+      "Identity" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.expandDims" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-16",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "Decoder",
+    "method" : "predict"
+  }
+]

Decoder.mlmodelc/model.mil

@@ -0,0 +1,54 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1025, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_1_dtype_0 = const()[name = tensor<string, []>("cast_1_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_9")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = targets_to_int16)[name = tensor<string, []>("cast_8")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(1025)];
+            tensor<int32, [1, 1]> add_1 = add(x = cast_1, y = slice_by_index_0)[name = tensor<string, []>("add_1")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_1, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("y_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_6")];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = h_in_to_fp16)[name = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = c_in_to_fp16)[name = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1312128)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4588992)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7865856)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = tensor<string, []>("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_0_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> obj_3_axes_0 = const()[name = tensor<string, []>("obj_3_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_3_cast_fp16 = expand_dims(axes = obj_3_axes_0, x = input_cast_fp16_1)[name = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> obj_axes_0 = const()[name = tensor<string, []>("obj_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 1, 640]> obj_cast_fp16 = expand_dims(axes = obj_axes_0, x = input_cast_fp16_2)[name = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [1, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [1, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}

Decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd90b58597ee2c172c672dffe13b1110898ba07394c1a15efc96cc8c6b18411b
+size 7871040

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9ef589f858f577a3e0d3df097ede3efe407d7152d3022a965a7e78744f84404
+size 7973

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd90b58597ee2c172c672dffe13b1110898ba07394c1a15efc96cc8c6b18411b
+size 7871040

Decoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "5F11CFA7-5015-4466-A6AA-0AE65C954D95": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "D180DDE9-941D-455C-B839-91C0843D02DE": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "D180DDE9-941D-455C-B839-91C0843D02DE"
+}

JointDecision.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42f7757cc6603cf025dd490169edef879c355b62a178243e1ca4d538e762af2f
+size 243

JointDecision.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ef957d12add316f594b2e8766202ce2b06341e3e5ebaea35b9ff2c3d7a0f4f7
+size 583

JointDecision.mlmodelc/metadata.json

@@ -0,0 +1,124 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet TDT-CTC single-step joint decision",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_id",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "token_prob",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1]",
+        "name" : "duration",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 64]",
+        "name" : "top_k_logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.reduceArgmax" : 2,
+      "Ios17.linear" : 3,
+      "Ios17.transpose" : 2,
+      "Ios17.sliceByIndex" : 2,
+      "Ios17.add" : 1,
+      "Ios17.topk" : 1,
+      "Ios16.relu" : 1,
+      "Ios16.softmax" : 1,
+      "Ios17.expandDims" : 3,
+      "Ios17.squeeze" : 1,
+      "Ios17.cast" : 6,
+      "Ios17.gatherAlongAxis" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1]",
+        "name" : "encoder_step",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_step",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-16",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "JointDecisionSingleStep",
+    "method" : "predict"
+  }
+]

JointDecision.mlmodelc/model.mil

@@ -0,0 +1,69 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder_step, tensor<fp32, [1, 512, 1]> encoder_step) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> encoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("encoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> decoder_step_to_fp16_dtype_0 = const()[name = tensor<string, []>("decoder_step_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [640, 512]> joint_module_enc_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(655488)))];
+            tensor<fp16, [1, 512, 1]> encoder_step_to_fp16 = cast(dtype = encoder_step_to_fp16_dtype_0, x = encoder_step)[name = tensor<string, []>("cast_5")];
+            tensor<fp16, [1, 1, 512]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_step_to_fp16)[name = tensor<string, []>("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(656832)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1476096)))];
+            tensor<fp16, [1, 640, 1]> decoder_step_to_fp16 = cast(dtype = decoder_step_to_fp16_dtype_0, x = decoder_step)[name = tensor<string, []>("cast_4")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_step_to_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = tensor<string, []>("op_23_cast_fp16")];
+            tensor<int32, [1]> var_24_axes_0 = const()[name = tensor<string, []>("op_24_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_24_cast_fp16 = expand_dims(axes = var_24_axes_0, x = linear_1_cast_fp16)[name = tensor<string, []>("op_24_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_24_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1030, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [1030, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1477440)))];
+            tensor<fp16, [1030]> joint_module_joint_net_2_bias_to_fp16 = const()[name = tensor<string, []>("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [1030]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2795904)))];
+            tensor<fp16, [1, 1, 1, 1030]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [4]> token_logits_begin_0 = const()[name = tensor<string, []>("token_logits_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> token_logits_end_0 = const()[name = tensor<string, []>("token_logits_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1025])];
+            tensor<bool, [4]> token_logits_end_mask_0 = const()[name = tensor<string, []>("token_logits_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 1, 1, 1025]> token_logits_cast_fp16 = slice_by_index(begin = token_logits_begin_0, end = token_logits_end_0, end_mask = token_logits_end_mask_0, x = linear_2_cast_fp16)[name = tensor<string, []>("token_logits_cast_fp16")];
+            tensor<int32, [4]> duration_logits_begin_0 = const()[name = tensor<string, []>("duration_logits_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 1025])];
+            tensor<int32, [4]> duration_logits_end_0 = const()[name = tensor<string, []>("duration_logits_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1030])];
+            tensor<bool, [4]> duration_logits_end_mask_0 = const()[name = tensor<string, []>("duration_logits_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 1, 1, 5]> duration_logits_cast_fp16 = slice_by_index(begin = duration_logits_begin_0, end = duration_logits_end_0, end_mask = duration_logits_end_mask_0, x = linear_2_cast_fp16)[name = tensor<string, []>("duration_logits_cast_fp16")];
+            tensor<int32, []> var_43_axis_0 = const()[name = tensor<string, []>("op_43_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_43_keep_dims_0 = const()[name = tensor<string, []>("op_43_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_43_output_dtype_0 = const()[name = tensor<string, []>("op_43_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> token_id = reduce_argmax(axis = var_43_axis_0, keep_dims = var_43_keep_dims_0, output_dtype = var_43_output_dtype_0, x = token_logits_cast_fp16)[name = tensor<string, []>("op_43_cast_fp16")];
+            tensor<int32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 1, 1, 1025]> token_probs_all_cast_fp16 = softmax(axis = var_49, x = token_logits_cast_fp16)[name = tensor<string, []>("token_probs_all_cast_fp16")];
+            tensor<int32, [1]> var_58_axes_0 = const()[name = tensor<string, []>("op_58_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1, 1, 1]> var_58 = expand_dims(axes = var_58_axes_0, x = token_id)[name = tensor<string, []>("op_58")];
+            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_61_validate_indices_0 = const()[name = tensor<string, []>("op_61_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_58_to_int16_dtype_0 = const()[name = tensor<string, []>("op_58_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1, 1, 1]> var_58_to_int16 = cast(dtype = var_58_to_int16_dtype_0, x = var_58)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1, 1, 1, 1]> var_61_cast_fp16_cast_int16 = gather_along_axis(axis = var_59, indices = var_58_to_int16, validate_indices = var_61_validate_indices_0, x = token_probs_all_cast_fp16)[name = tensor<string, []>("op_61_cast_fp16_cast_int16")];
+            tensor<int32, [1]> var_63_axes_0 = const()[name = tensor<string, []>("op_63_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1, 1]> var_63_cast_fp16 = squeeze(axes = var_63_axes_0, x = var_61_cast_fp16_cast_int16)[name = tensor<string, []>("op_63_cast_fp16")];
+            tensor<string, []> var_63_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_63_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> var_66_axis_0 = const()[name = tensor<string, []>("op_66_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_66_keep_dims_0 = const()[name = tensor<string, []>("op_66_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<string, []> var_66_output_dtype_0 = const()[name = tensor<string, []>("op_66_output_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 1, 1]> duration = reduce_argmax(axis = var_66_axis_0, keep_dims = var_66_keep_dims_0, output_dtype = var_66_output_dtype_0, x = duration_logits_cast_fp16)[name = tensor<string, []>("op_66_cast_fp16")];
+            tensor<int32, []> var_72 = const()[name = tensor<string, []>("op_72"), val = tensor<int32, []>(64)];
+            tensor<int32, []> var_76_axis_0 = const()[name = tensor<string, []>("op_76_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<bool, []> var_76_ascending_0 = const()[name = tensor<string, []>("op_76_ascending_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> var_76_sort_0 = const()[name = tensor<string, []>("op_76_sort_0"), val = tensor<bool, []>(true)];
+            tensor<bool, []> var_76_return_indices_0 = const()[name = tensor<string, []>("op_76_return_indices_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> var_76_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_cast_int16_output_indices_dtype_0"), val = tensor<string, []>("uint16")];
+            tensor<fp16, [1, 1, 1, 64]> var_76_cast_fp16_cast_int16_0, tensor<uint16, [1, 1, 1, 64]> var_76_cast_fp16_cast_int16_1 = topk(ascending = var_76_ascending_0, axis = var_76_axis_0, k = var_72, output_indices_dtype = var_76_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_76_return_indices_0, sort = var_76_sort_0, x = token_logits_cast_fp16)[name = tensor<string, []>("op_76_cast_fp16_cast_int16")];
+            tensor<string, []> var_76_cast_fp16_cast_int16_1_to_int32_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_cast_int16_1_to_int32_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_76_cast_fp16_0_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_76_cast_fp16_0_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, 64]> top_k_logits = cast(dtype = var_76_cast_fp16_0_to_fp32_dtype_0, x = var_76_cast_fp16_cast_int16_0)[name = tensor<string, []>("cast_0")];
+            tensor<int32, [1, 1, 1, 64]> top_k_ids = cast(dtype = var_76_cast_fp16_cast_int16_1_to_int32_dtype_0, x = var_76_cast_fp16_cast_int16_1)[name = tensor<string, []>("cast_1")];
+            tensor<fp32, [1, 1, 1]> token_prob = cast(dtype = var_63_cast_fp16_to_fp32_dtype_0, x = var_63_cast_fp16)[name = tensor<string, []>("cast_2")];
+        } -> (token_id, token_prob, duration, top_k_ids, top_k_logits);
+}

JointDecision.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3f771cb65b190f1873e39629676ed79b65a8361522f451b37bdba8b1106e6ff
+size 2798028

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b5bf959656c8f51446b19664a8ded1a1a8630ddd42a75325387e3c9dd6400d4
+size 10668

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3f771cb65b190f1873e39629676ed79b65a8361522f451b37bdba8b1106e6ff
+size 2798028

JointDecisionSingleStep.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "9837468A-52AB-4172-B88E-85F05A9261A0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "E9AA441F-A636-4AF4-BE59-ECD44D3C22F6": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "9837468A-52AB-4172-B88E-85F05A9261A0"
+}

Preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d813d089b82b1edfdc558d902d91421c527c2db4ba85e36c4a611ea99eba8248
+size 243

Preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69e29c6cf43349318661a4aaade4bc0cc124a5780582d283614056cc76eb2907
+size 499

Preprocessor.mlmodelc/metadata.json

@@ -0,0 +1,116 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet TDT-CTC Mel+Encoder (15s window)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512 × 188)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 188]",
+        "name" : "encoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoder_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Tile" : 2,
+      "Stack" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios16.silu" : 51,
+      "Ios17.sub" : 4,
+      "Identity" : 1,
+      "Ios16.reduceSum" : 4,
+      "Pad" : 35,
+      "Ios17.logicalNot" : 2,
+      "Ios17.sliceByIndex" : 37,
+      "Ios17.concat" : 1,
+      "Ios17.pow" : 2,
+      "Ios17.sqrt" : 1,
+      "Ios17.floor" : 3,
+      "Ios17.floorDiv" : 1,
+      "Ios17.expandDims" : 13,
+      "Ios17.less" : 2,
+      "Ios17.mul" : 72,
+      "Ios17.transpose" : 123,
+      "Ios17.matmul" : 52,
+      "Ios16.sigmoid" : 17,
+      "Ios17.conv" : 58,
+      "Ios17.reshape" : 105,
+      "Split" : 17,
+      "Ios17.log" : 1,
+      "Ios17.cast" : 7,
+      "Ios17.linear" : 137,
+      "Ios17.realDiv" : 3,
+      "Select" : 54,
+      "Ios16.relu" : 3,
+      "Ios16.softmax" : 17,
+      "Ios17.add" : 129,
+      "Ios17.layerNorm" : 85,
+      "Ios17.logicalAnd" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 240000)",
+        "shortDescription" : "",
+        "shape" : "[1, 240000]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-16",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "Preprocessor",
+    "method" : "predict"
+  }
+]

Preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1c90b88b52667ee4f349b39994003cc2a9fb1b12c310752399d90c880286a7b
+size 215951360

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cdb59bc9293df8a1d993f534033e212d64fb12bdfad98fe96964b804b8d4d1e
+size 525105

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1c90b88b52667ee4f349b39994003cc2a9fb1b12c310752399d90c880286a7b
+size 215951360

Preprocessor.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "39C01B3A-5AD3-4DB9-950C-0AFE37EBAB79": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "A9DC3D62-1D15-47F5-86AD-D33E82326E53": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "A9DC3D62-1D15-47F5-86AD-D33E82326E53"
+}

README.md

@@ -0,0 +1,89 @@
+---
+license: cc-by-4.0
+language:
+- en
+metrics:
+- wer
+base_model:
+- nvidia/parakeet-tdt_ctc-110m
+pipeline_tag: automatic-speech-recognition
+tags:
+- automatic-speech-recognition
+- speech
+- audio
+- Transducer
+- TDT
+- FastConformer
+- Conformer
+- pytorch
+- NeMo
+- hf-asr-leaderboard
+---
+# Parakeet-TDT-CTC 110M — CoreML
+
+CoreML export of [nvidia/parakeet-tdt_ctc-110m](https://huggingface.co/nvidia/parakeet-tdt_ctc-110m) for on-device speech recognition on Apple Silicon via [FluidAudio](https://github.com/FluidInference/FluidAudio).
+
+
+## CoreML Components
+
+| File | Size | Description |
+|------|------|-------------|
+| `Preprocessor.mlmodelc` | 207 MB | Fused mel-spectrogram + FastConformer encoder |
+| `Decoder.mlmodelc` | 7.5 MB | 1-layer LSTM prediction network |
+| `JointDecision.mlmodelc` | 2.7 MB | Single-step joint network (token + duration) |
+| `parakeet_vocab.json` | 18 KB | 1024-token BPE vocabulary |
+| `config.json` | 2.5 KB | Model metadata and I/O contracts |
+
+**Input:** 16 kHz mono audio, fixed 15-second window (240,000 samples).
+**Output:** Token IDs, probabilities, and TDT duration predictions per encoder frame.
+
+## Performance
+
+Benchmarked with FluidAudio CLI on Apple M2 (release build):
+
+| Benchmark | WER |
+|-----------|-----|
+| LibriSpeech test-clean | **3.0%** |
+| RTFx (overall) | **102x** real-time |
+| Peak memory | 0.3 GB |
+
+NVIDIA's reference WER (greedy, GPU):
+
+| Benchmark | WER |
+|-----------|-----|
+| LibriSpeech test-clean | 2.4% |
+| LibriSpeech test-other | 5.2% |
+| AMI | 15.88% |
+| Earnings-22 | 12.42% |
+| GigaSpeech | 10.52% |
+| TEDLIUM-v3 | 4.16% |
+
+## Usage with FluidAudio
+
+```bash
+# Transcribe
+fluidaudiocli transcribe audio.wav --model-version tdt-ctc-110m
+
+# Benchmark
+fluidaudiocli asr-benchmark --subset test-clean --model-version tdt-ctc-110m
+```
+
+Models auto-download from this repo on first use. To pre-fetch:
+
+```bash
+fluidaudiocli download --model-version tdt-ctc-110m
+```
+
+## Conversion
+
+Exported from NeMo using [mobius/models/stt/parakeet-tdt-ctc-110m/coreml/convert-tdt-coreml.py](https://github.com/FluidInference/mobius):
+
+- Preprocessor fuses mel-spectrogram extraction and the FastConformer encoder into a single CoreML model
+- JointDecision is the single-step variant (encoder_step + decoder_step inputs) used by FluidAudio's TDT decoder
+- All models exported as MLProgram (iOS 17+ / macOS 14+), float32 precision
+
+## References
+
+- [Fast Conformer with Linearly Scalable Attention](https://arxiv.org/abs/2305.05084)
+- [Efficient Sequence Transduction by Jointly Predicting Tokens and Durations](https://arxiv.org/abs/2304.06795)
+- [NVIDIA NeMo](https://github.com/NVIDIA/NeMo)

config.json

@@ -0,0 +1,147 @@
+{
+  "model_id": "nvidia/parakeet-tdt_ctc-110m",
+  "model_type": "hybrid_tdt_ctc",
+  "sample_rate": 16000,
+  "max_audio_seconds": 15.0,
+  "max_audio_samples": 240000,
+  "max_symbol_steps": 1,
+  "vocab_size": 1024,
+  "blank_id": 1024,
+  "joint_extra_outputs": 5,
+  "duration_bins": [
+    0,
+    1,
+    2,
+    3,
+    4
+  ],
+  "encoder_dim": 512,
+  "decoder_dim": 640,
+  "decoder_hidden": 640,
+  "decoder_layers": 1,
+  "checkpoint": {
+    "type": "pretrained",
+    "model_id": "nvidia/parakeet-tdt_ctc-110m"
+  },
+  "coreml": {
+    "compute_units": "CPU_ONLY",
+    "compute_precision": "FLOAT32"
+  },
+  "components": {
+    "preprocessor": {
+      "inputs": {
+        "audio_signal": [
+          1,
+          240000
+        ],
+        "audio_length": [
+          1
+        ]
+      },
+      "outputs": {
+        "encoder": [
+          1,
+          512,
+          188
+        ],
+        "encoder_length": [
+          1
+        ]
+      },
+      "path": "Preprocessor.mlpackage"
+    },
+    "decoder": {
+      "inputs": {
+        "targets": [
+          1,
+          1
+        ],
+        "target_length": [
+          1
+        ],
+        "h_in": [
+          1,
+          1,
+          640
+        ],
+        "c_in": [
+          1,
+          1,
+          640
+        ]
+      },
+      "outputs": {
+        "decoder": [
+          1,
+          640,
+          1
+        ],
+        "h_out": [
+          1,
+          1,
+          640
+        ],
+        "c_out": [
+          1,
+          1,
+          640
+        ]
+      },
+      "path": "Decoder.mlpackage"
+    },
+    "joint_decision": {
+      "inputs": {
+        "encoder": [
+          1,
+          512,
+          188
+        ],
+        "decoder": [
+          1,
+          640,
+          1
+        ]
+      },
+      "outputs": {
+        "token_id": "int32",
+        "token_prob": "float32",
+        "duration": "int32"
+      },
+      "path": "JointDecision.mlpackage"
+    },
+    "joint_decision_single_step": {
+      "inputs": {
+        "encoder_step": [
+          1,
+          512,
+          1
+        ],
+        "decoder_step": [
+          1,
+          640,
+          1
+        ]
+      },
+      "outputs": {
+        "token_id": [
+          1,
+          1,
+          1
+        ],
+        "token_prob": [
+          1,
+          1,
+          1
+        ],
+        "duration": [
+          1,
+          1,
+          1
+        ],
+        "top_k_ids": "int32",
+        "top_k_logits": "float32"
+      },
+      "path": "JointDecisionSingleStep.mlpackage"
+    }
+  }
+}

parakeet_vocab.json

@@ -0,0 +1,1026 @@
+{
+  "0": "<unk>",
+  "1": "▁t",
+  "2": "▁th",
+  "3": "▁a",
+  "4": "in",
+  "5": "re",
+  "6": "▁the",
+  "7": "▁w",
+  "8": "▁s",
+  "9": "▁o",
+  "10": "er",
+  "11": "ou",
+  "12": "at",
+  "13": "nd",
+  "14": "it",
+  "15": "▁h",
+  "16": "▁c",
+  "17": "▁b",
+  "18": "is",
+  "19": "en",
+  "20": "on",
+  "21": "ing",
+  "22": "▁f",
+  "23": "▁to",
+  "24": "▁m",
+  "25": "es",
+  "26": "▁p",
+  "27": "or",
+  "28": "an",
+  "29": "▁d",
+  "30": "ll",
+  "31": "▁I",
+  "32": "ed",
+  "33": "▁and",
+  "34": "▁l",
+  "35": "▁of",
+  "36": "▁in",
+  "37": "▁y",
+  "38": "ar",
+  "39": "▁g",
+  "40": "▁you",
+  "41": "as",
+  "42": "om",
+  "43": "▁n",
+  "44": "ve",
+  "45": "▁that",
+  "46": "le",
+  "47": "ic",
+  "48": "us",
+  "49": "ow",
+  "50": "et",
+  "51": "al",
+  "52": "▁e",
+  "53": "ut",
+  "54": "▁it",
+  "55": "ot",
+  "56": "▁be",
+  "57": "▁T",
+  "58": "ion",
+  "59": "▁is",
+  "60": "▁wh",
+  "61": "▁re",
+  "62": "▁on",
+  "63": "▁we",
+  "64": "ent",
+  "65": "▁A",
+  "66": "ay",
+  "67": "▁ha",
+  "68": "▁Th",
+  "69": "id",
+  "70": "▁S",
+  "71": "ac",
+  "72": "gh",
+  "73": "ver",
+  "74": "ke",
+  "75": "▁for",
+  "76": "im",
+  "77": "ly",
+  "78": "ur",
+  "79": "ld",
+  "80": "▁he",
+  "81": "▁st",
+  "82": "all",
+  "83": "ro",
+  "84": "st",
+  "85": "se",
+  "86": "ct",
+  "87": "ith",
+  "88": "ir",
+  "89": "am",
+  "90": "▁this",
+  "91": "if",
+  "92": "▁W",
+  "93": "oo",
+  "94": "ri",
+  "95": "▁was",
+  "96": "ght",
+  "97": "▁u",
+  "98": "▁with",
+  "99": "ad",
+  "100": "ch",
+  "101": "▁se",
+  "102": "▁k",
+  "103": "▁an",
+  "104": "▁The",
+  "105": "▁li",
+  "106": "▁do",
+  "107": "▁B",
+  "108": "▁have",
+  "109": "▁as",
+  "110": "th",
+  "111": "▁are",
+  "112": "▁sh",
+  "113": "ust",
+  "114": "ce",
+  "115": "ally",
+  "116": "ill",
+  "117": "▁H",
+  "118": "▁j",
+  "119": "ter",
+  "120": "▁go",
+  "121": "▁And",
+  "122": "ation",
+  "123": "▁C",
+  "124": "▁so",
+  "125": "ome",
+  "126": "▁not",
+  "127": "op",
+  "128": "il",
+  "129": "ore",
+  "130": "▁ne",
+  "131": "▁can",
+  "132": "▁me",
+  "133": "▁at",
+  "134": "ould",
+  "135": "ant",
+  "136": "▁M",
+  "137": "▁like",
+  "138": "ere",
+  "139": "▁they",
+  "140": "ra",
+  "141": "ers",
+  "142": "▁ab",
+  "143": "▁de",
+  "144": "▁kn",
+  "145": "ge",
+  "146": "▁Y",
+  "147": "▁ch",
+  "148": "ul",
+  "149": "pp",
+  "150": "▁or",
+  "151": "▁al",
+  "152": "▁con",
+  "153": "▁com",
+  "154": "ess",
+  "155": "▁su",
+  "156": "out",
+  "157": "▁your",
+  "158": "▁So",
+  "159": "ate",
+  "160": "▁one",
+  "161": "▁all",
+  "162": "▁ex",
+  "163": "est",
+  "164": "▁fr",
+  "165": "▁just",
+  "166": "▁pro",
+  "167": "▁know",
+  "168": "▁O",
+  "169": "ain",
+  "170": "▁but",
+  "171": "ol",
+  "172": "ive",
+  "173": "▁v",
+  "174": "use",
+  "175": "very",
+  "176": "art",
+  "177": "qu",
+  "178": "▁my",
+  "179": "el",
+  "180": "▁N",
+  "181": "nt",
+  "182": "▁It",
+  "183": "▁what",
+  "184": "ab",
+  "185": "▁P",
+  "186": "▁wor",
+  "187": "▁out",
+  "188": "▁there",
+  "189": "▁up",
+  "190": "um",
+  "191": "▁from",
+  "192": "pe",
+  "193": "▁tw",
+  "194": "▁r",
+  "195": "and",
+  "196": "ight",
+  "197": "ort",
+  "198": "un",
+  "199": "▁L",
+  "200": "ist",
+  "201": "▁about",
+  "202": "ide",
+  "203": "ig",
+  "204": "ake",
+  "205": "▁D",
+  "206": "em",
+  "207": "os",
+  "208": "king",
+  "209": "rou",
+  "210": "ind",
+  "211": "our",
+  "212": "res",
+  "213": "▁We",
+  "214": "▁get",
+  "215": "▁E",
+  "216": "▁G",
+  "217": "ack",
+  "218": "▁le",
+  "219": "ity",
+  "220": "od",
+  "221": "▁F",
+  "222": "ard",
+  "223": "▁pl",
+  "224": "▁our",
+  "225": "▁int",
+  "226": "ment",
+  "227": "▁will",
+  "228": "ies",
+  "229": "▁by",
+  "230": "ink",
+  "231": "ca",
+  "232": "▁if",
+  "233": "red",
+  "234": "her",
+  "235": "ie",
+  "236": "▁us",
+  "237": "▁some",
+  "238": "▁don",
+  "239": "ven",
+  "240": "ood",
+  "241": "ast",
+  "242": "▁R",
+  "243": "▁his",
+  "244": "▁tim",
+  "245": "▁tr",
+  "246": "▁more",
+  "247": "ich",
+  "248": "ous",
+  "249": "ame",
+  "250": "▁going",
+  "251": "▁had",
+  "252": "▁them",
+  "253": "ook",
+  "254": "▁pe",
+  "255": "▁Wh",
+  "256": "▁You",
+  "257": "▁But",
+  "258": "ine",
+  "259": "▁here",
+  "260": "▁would",
+  "261": "cause",
+  "262": "right",
+  "263": "so",
+  "264": "ost",
+  "265": "ure",
+  "266": "▁has",
+  "267": "ect",
+  "268": "▁think",
+  "269": "▁fe",
+  "270": "ong",
+  "271": "▁see",
+  "272": "▁when",
+  "273": "▁who",
+  "274": "▁were",
+  "275": "▁really",
+  "276": "▁their",
+  "277": "▁want",
+  "278": "one",
+  "279": "ople",
+  "280": "▁then",
+  "281": "▁time",
+  "282": "▁sa",
+  "283": "ap",
+  "284": "▁te",
+  "285": "▁He",
+  "286": "▁ye",
+  "287": "ck",
+  "288": "▁her",
+  "289": "▁thing",
+  "290": "▁right",
+  "291": "▁which",
+  "292": "itt",
+  "293": "ice",
+  "294": "act",
+  "295": "▁people",
+  "296": "ty",
+  "297": "▁two",
+  "298": "▁J",
+  "299": "▁im",
+  "300": "ther",
+  "301": "ci",
+  "302": "ose",
+  "303": "▁cl",
+  "304": "▁qu",
+  "305": "▁man",
+  "306": "▁also",
+  "307": "ree",
+  "308": "▁en",
+  "309": "ud",
+  "310": "▁how",
+  "311": "reat",
+  "312": "ak",
+  "313": "hing",
+  "314": "ag",
+  "315": "▁any",
+  "316": "ff",
+  "317": "ace",
+  "318": "per",
+  "319": "▁because",
+  "320": "▁very",
+  "321": "own",
+  "322": "▁ad",
+  "323": "▁act",
+  "324": "▁been",
+  "325": "▁now",
+  "326": "▁ag",
+  "327": "▁into",
+  "328": "▁comp",
+  "329": "ars",
+  "330": "ions",
+  "331": "are",
+  "332": "ite",
+  "333": "iv",
+  "334": "▁these",
+  "335": "ays",
+  "336": "ep",
+  "337": "▁This",
+  "338": "▁she",
+  "339": "ans",
+  "340": "ah",
+  "341": "een",
+  "342": "▁over",
+  "343": "ry",
+  "344": "▁lo",
+  "345": "age",
+  "346": "▁pr",
+  "347": "▁sp",
+  "348": "ue",
+  "349": "▁co",
+  "350": "ick",
+  "351": "ber",
+  "352": "▁did",
+  "353": "ip",
+  "354": "ach",
+  "355": "▁back",
+  "356": "▁no",
+  "357": "▁cont",
+  "358": "▁other",
+  "359": "▁every",
+  "360": "pt",
+  "361": "▁need",
+  "362": "▁him",
+  "363": "▁U",
+  "364": "▁In",
+  "365": "▁work",
+  "366": "irst",
+  "367": "▁part",
+  "368": "▁look",
+  "369": "ittle",
+  "370": "ble",
+  "371": "iz",
+  "372": "▁un",
+  "373": "▁make",
+  "374": "omet",
+  "375": "nder",
+  "376": "ish",
+  "377": "na",
+  "378": "▁little",
+  "379": "▁off",
+  "380": "▁than",
+  "381": "▁got",
+  "382": "ually",
+  "383": "▁per",
+  "384": "▁good",
+  "385": "▁way",
+  "386": "▁could",
+  "387": "▁ac",
+  "388": "▁imp",
+  "389": "able",
+  "390": "▁where",
+  "391": "iff",
+  "392": "▁That",
+  "393": "▁res",
+  "394": "ount",
+  "395": "pl",
+  "396": "ance",
+  "397": "▁first",
+  "398": "▁ro",
+  "399": "▁pre",
+  "400": "ass",
+  "401": "▁say",
+  "402": "int",
+  "403": "ated",
+  "404": "ire",
+  "405": "uch",
+  "406": "ase",
+  "407": "▁somet",
+  "408": "ound",
+  "409": "▁down",
+  "410": "▁diff",
+  "411": "sel",
+  "412": "▁gu",
+  "413": "▁am",
+  "414": "ress",
+  "415": "▁lot",
+  "416": "ence",
+  "417": "▁dis",
+  "418": "orm",
+  "419": "ix",
+  "420": "▁po",
+  "421": "ving",
+  "422": "enty",
+  "423": "▁K",
+  "424": "▁spe",
+  "425": "und",
+  "426": "he",
+  "427": "▁much",
+  "428": "▁ar",
+  "429": "round",
+  "430": "▁app",
+  "431": "co",
+  "432": "ark",
+  "433": "▁new",
+  "434": "ater",
+  "435": "ult",
+  "436": "end",
+  "437": "▁even",
+  "438": "▁start",
+  "439": "ations",
+  "440": "rough",
+  "441": "ile",
+  "442": "fter",
+  "443": "▁well",
+  "444": "be",
+  "445": "▁They",
+  "446": "▁three",
+  "447": "ign",
+  "448": "ild",
+  "449": "▁said",
+  "450": "ough",
+  "451": "ang",
+  "452": "▁too",
+  "453": "ade",
+  "454": "▁bl",
+  "455": "ens",
+  "456": "▁inc",
+  "457": "ia",
+  "458": "▁those",
+  "459": "▁mo",
+  "460": "▁take",
+  "461": "▁through",
+  "462": "▁fl",
+  "463": "▁kind",
+  "464": "▁things",
+  "465": "▁bet",
+  "466": "▁only",
+  "467": "▁St",
+  "468": "▁let",
+  "469": "cess",
+  "470": "▁Ch",
+  "471": "ary",
+  "472": "vel",
+  "473": "▁If",
+  "474": "xt",
+  "475": "other",
+  "476": "av",
+  "477": "ical",
+  "478": "ord",
+  "479": "▁again",
+  "480": "▁something",
+  "481": "onna",
+  "482": "fore",
+  "483": "▁may",
+  "484": "ting",
+  "485": "▁bu",
+  "486": "▁differe",
+  "487": "urn",
+  "488": "▁gonna",
+  "489": "▁does",
+  "490": "uct",
+  "491": "og",
+  "492": "▁twenty",
+  "493": "▁gr",
+  "494": "▁Ye",
+  "495": "wn",
+  "496": "▁should",
+  "497": "▁comm",
+  "498": "ition",
+  "499": "▁under",
+  "500": "▁hel",
+  "501": "ory",
+  "502": "▁fo",
+  "503": "▁use",
+  "504": "igh",
+  "505": "ife",
+  "506": "▁actually",
+  "507": "▁tal",
+  "508": "▁call",
+  "509": "ents",
+  "510": "ious",
+  "511": "ull",
+  "512": "▁There",
+  "513": "▁Yeah",
+  "514": "▁most",
+  "515": "▁ke",
+  "516": "ors",
+  "517": "ved",
+  "518": "ys",
+  "519": "▁sc",
+  "520": "▁happ",
+  "521": "ope",
+  "522": "▁help",
+  "523": "atch",
+  "524": "▁What",
+  "525": "▁rem",
+  "526": "ple",
+  "527": "▁Now",
+  "528": "▁br",
+  "529": "ool",
+  "530": "oth",
+  "531": "▁four",
+  "532": "self",
+  "533": "▁str",
+  "534": "ne",
+  "535": "thing",
+  "536": "▁put",
+  "537": "ial",
+  "538": "▁great",
+  "539": "ail",
+  "540": "ub",
+  "541": "ning",
+  "542": "▁sm",
+  "543": "▁feel",
+  "544": "▁five",
+  "545": "ody",
+  "546": "undred",
+  "547": "iss",
+  "548": "ank",
+  "549": "get",
+  "550": "aking",
+  "551": "▁many",
+  "552": "▁hundred",
+  "553": "▁years",
+  "554": "▁being",
+  "555": "▁come",
+  "556": "▁mean",
+  "557": "ily",
+  "558": "▁different",
+  "559": "▁after",
+  "560": "▁ser",
+  "561": "▁show",
+  "562": "form",
+  "563": "ful",
+  "564": "oy",
+  "565": "▁six",
+  "566": "▁vide",
+  "567": "▁V",
+  "568": "▁its",
+  "569": "▁point",
+  "570": "▁day",
+  "571": "▁des",
+  "572": "ons",
+  "573": "▁bit",
+  "574": "▁bel",
+  "575": "▁before",
+  "576": "▁aw",
+  "577": "▁end",
+  "578": "▁Oh",
+  "579": "▁still",
+  "580": "ath",
+  "581": "▁long",
+  "582": "▁'",
+  "583": "ise",
+  "584": "ob",
+  "585": "day",
+  "586": "▁add",
+  "587": "ft",
+  "588": "ves",
+  "589": "ces",
+  "590": "ady",
+  "591": "▁cr",
+  "592": "▁around",
+  "593": "▁try",
+  "594": "les",
+  "595": "vers",
+  "596": "kay",
+  "597": "ian",
+  "598": "ates",
+  "599": "▁find",
+  "600": "ward",
+  "601": "▁As",
+  "602": "▁eight",
+  "603": "lic",
+  "604": "▁same",
+  "605": "▁pos",
+  "606": "▁em",
+  "607": "▁made",
+  "608": "▁supp",
+  "609": "▁life",
+  "610": "▁Be",
+  "611": "pect",
+  "612": "▁dec",
+  "613": "▁play",
+  "614": "ange",
+  "615": "▁att",
+  "616": "▁pers",
+  "617": "ways",
+  "618": "▁high",
+  "619": "▁hand",
+  "620": "▁next",
+  "621": "▁cons",
+  "622": "▁own",
+  "623": "▁inv",
+  "624": "ower",
+  "625": "▁ind",
+  "626": "ert",
+  "627": "ng",
+  "628": "ave",
+  "629": "▁year",
+  "630": "▁big",
+  "631": "ating",
+  "632": "▁world",
+  "633": "▁rel",
+  "634": "▁sure",
+  "635": "▁tra",
+  "636": "ew",
+  "637": "ered",
+  "638": "▁fin",
+  "639": "▁Well",
+  "640": "▁sl",
+  "641": "▁doing",
+  "642": "bs",
+  "643": "▁set",
+  "644": "▁rec",
+  "645": "ual",
+  "646": "cial",
+  "647": "▁ph",
+  "648": "erm",
+  "649": "▁love",
+  "650": "ph",
+  "651": "▁real",
+  "652": "▁last",
+  "653": "ict",
+  "654": "▁bo",
+  "655": "▁ra",
+  "656": "ible",
+  "657": "▁wr",
+  "658": "mer",
+  "659": "▁count",
+  "660": "ities",
+  "661": "▁always",
+  "662": "inet",
+  "663": "ments",
+  "664": "uc",
+  "665": "▁might",
+  "666": "▁inter",
+  "667": "▁video",
+  "668": "gin",
+  "669": "▁tell",
+  "670": "▁never",
+  "671": "vent",
+  "672": "▁import",
+  "673": "ied",
+  "674": "▁sy",
+  "675": "▁How",
+  "676": "ically",
+  "677": "ought",
+  "678": "▁thir",
+  "679": "▁rep",
+  "680": "ks",
+  "681": "ib",
+  "682": "▁fam",
+  "683": "ject",
+  "684": "▁bas",
+  "685": "▁She",
+  "686": "▁give",
+  "687": "akes",
+  "688": "▁ninet",
+  "689": "▁reg",
+  "690": "▁min",
+  "691": "▁op",
+  "692": "▁def",
+  "693": "▁didn",
+  "694": "te",
+  "695": "▁cour",
+  "696": "▁why",
+  "697": "▁ent",
+  "698": "▁place",
+  "699": "▁ins",
+  "700": "▁car",
+  "701": "ather",
+  "702": "▁person",
+  "703": "ular",
+  "704": "▁inst",
+  "705": "▁prod",
+  "706": "lect",
+  "707": "▁Al",
+  "708": "▁today",
+  "709": "▁bec",
+  "710": "▁sur",
+  "711": "▁All",
+  "712": "▁another",
+  "713": "▁bus",
+  "714": "▁keep",
+  "715": "ell",
+  "716": "ese",
+  "717": "riend",
+  "718": "▁quest",
+  "719": "▁talk",
+  "720": "als",
+  "721": "ings",
+  "722": "▁mon",
+  "723": "cond",
+  "724": "old",
+  "725": "▁acc",
+  "726": "▁la",
+  "727": "▁num",
+  "728": "ident",
+  "729": "▁che",
+  "730": "iness",
+  "731": "▁turn",
+  "732": "▁ear",
+  "733": "▁No",
+  "734": "ousand",
+  "735": "▁better",
+  "736": "ific",
+  "737": "▁loo",
+  "738": "▁gl",
+  "739": "oc",
+  "740": "▁important",
+  "741": "ited",
+  "742": "▁An",
+  "743": "▁thousand",
+  "744": "ility",
+  "745": "llow",
+  "746": "▁used",
+  "747": "▁gen",
+  "748": "▁sim",
+  "749": "li",
+  "750": "▁happen",
+  "751": "▁Un",
+  "752": "▁Let",
+  "753": "air",
+  "754": "ock",
+  "755": "ably",
+  "756": "gg",
+  "757": "▁watch",
+  "758": "▁For",
+  "759": "▁sw",
+  "760": "ren",
+  "761": "ute",
+  "762": "ever",
+  "763": "▁pol",
+  "764": "▁sch",
+  "765": "▁When",
+  "766": "▁such",
+  "767": "▁fif",
+  "768": "▁home",
+  "769": "▁cle",
+  "770": "▁contin",
+  "771": "ouse",
+  "772": "▁friend",
+  "773": "uring",
+  "774": "▁Okay",
+  "775": "gr",
+  "776": "▁able",
+  "777": "▁stud",
+  "778": "▁eff",
+  "779": "hip",
+  "780": "body",
+  "781": "▁top",
+  "782": "ness",
+  "783": "▁exper",
+  "784": "▁pret",
+  "785": "▁both",
+  "786": "▁done",
+  "787": "cri",
+  "788": "▁mark",
+  "789": "▁while",
+  "790": "▁old",
+  "791": "ros",
+  "792": "ont",
+  "793": "▁second",
+  "794": "ative",
+  "795": "▁thought",
+  "796": "▁best",
+  "797": "▁found",
+  "798": "iew",
+  "799": "▁belie",
+  "800": "▁each",
+  "801": "erest",
+  "802": "▁tri",
+  "803": "▁eas",
+  "804": "▁ca",
+  "805": "▁fact",
+  "806": "▁care",
+  "807": "▁fun",
+  "808": "atter",
+  "809": "ures",
+  "810": "▁head",
+  "811": "▁lear",
+  "812": "▁water",
+  "813": "▁hard",
+  "814": "▁few",
+  "815": "▁side",
+  "816": "ween",
+  "817": "▁exp",
+  "818": "▁away",
+  "819": "its",
+  "820": "▁ext",
+  "821": "lud",
+  "822": "▁run",
+  "823": "▁trans",
+  "824": "ince",
+  "825": "▁sk",
+  "826": "▁open",
+  "827": "cus",
+  "828": "▁between",
+  "829": "▁called",
+  "830": "▁wee",
+  "831": "▁pretty",
+  "832": "ason",
+  "833": "▁far",
+  "834": "ember",
+  "835": "omm",
+  "836": "▁interest",
+  "837": "any",
+  "838": "ner",
+  "839": "uff",
+  "840": "▁pres",
+  "841": "▁cur",
+  "842": "▁child",
+  "843": "ee",
+  "844": "▁toget",
+  "845": "▁together",
+  "846": "olog",
+  "847": "▁God",
+  "848": "ond",
+  "849": "▁char",
+  "850": "▁looking",
+  "851": "stem",
+  "852": "az",
+  "853": "cent",
+  "854": "▁ob",
+  "855": "▁ass",
+  "856": "land",
+  "857": "▁doesn",
+  "858": "▁business",
+  "859": "▁course",
+  "860": "▁ten",
+  "861": "ps",
+  "862": "arch",
+  "863": "ced",
+  "864": "ms",
+  "865": "ize",
+  "866": "nce",
+  "867": "▁ref",
+  "868": "▁name",
+  "869": "ross",
+  "870": "▁grow",
+  "871": "oney",
+  "872": "▁went",
+  "873": "ics",
+  "874": "teen",
+  "875": "▁cou",
+  "876": "▁prob",
+  "877": "▁ret",
+  "878": "▁guys",
+  "879": "▁came",
+  "880": "ash",
+  "881": "led",
+  "882": "▁Eur",
+  "883": "ues",
+  "884": "▁ide",
+  "885": "gan",
+  "886": "▁everything",
+  "887": "▁getting",
+  "888": "▁ask",
+  "889": "▁cor",
+  "890": "▁build",
+  "891": "▁sign",
+  "892": "▁small",
+  "893": "uck",
+  "894": "▁el",
+  "895": "▁col",
+  "896": "▁Is",
+  "897": "ational",
+  "898": "stand",
+  "899": "cy",
+  "900": "▁conf",
+  "901": "der",
+  "902": "▁bre",
+  "903": "▁cap",
+  "904": "▁mod",
+  "905": "ets",
+  "906": "ike",
+  "907": "▁number",
+  "908": "▁comple",
+  "909": "ertain",
+  "910": "▁ever",
+  "911": "▁coll",
+  "912": "▁hum",
+  "913": "▁Europe",
+  "914": "▁cre",
+  "915": "▁met",
+  "916": "▁exam",
+  "917": "▁move",
+  "918": "▁pass",
+  "919": "▁left",
+  "920": "▁system",
+  "921": "▁includ",
+  "922": "▁Thank",
+  "923": "cept",
+  "924": "▁wom",
+  "925": "▁product",
+  "926": "ten",
+  "927": "▁rest",
+  "928": "▁probably",
+  "929": "▁dri",
+  "930": "▁Do",
+  "931": "▁gener",
+  "932": "▁anything",
+  "933": "▁lar",
+  "934": "▁My",
+  "935": "▁school",
+  "936": "▁lead",
+  "937": "▁sub",
+  "938": "▁ty",
+  "939": "▁plan",
+  "940": "▁seem",
+  "941": "▁whole",
+  "942": "irect",
+  "943": "▁light",
+  "944": "▁must",
+  "945": "▁mom",
+  "946": "▁opp",
+  "947": "▁support",
+  "948": "▁family",
+  "949": "ices",
+  "950": "amp",
+  "951": "▁proble",
+  "952": "▁dr",
+  "953": "ready",
+  "954": "▁using",
+  "955": "ense",
+  "956": "▁prov",
+  "957": "ush",
+  "958": "ax",
+  "959": "▁power",
+  "960": "▁Re",
+  "961": "alth",
+  "962": "▁ev",
+  "963": "▁stand",
+  "964": "▁war",
+  "965": "ts",
+  "966": "▁",
+  "967": "e",
+  "968": "t",
+  "969": "o",
+  "970": "a",
+  "971": "n",
+  "972": "i",
+  "973": "s",
+  "974": "r",
+  "975": "h",
+  "976": "l",
+  "977": "d",
+  "978": "u",
+  "979": "c",
+  "980": "m",
+  "981": "y",
+  "982": "g",
+  "983": "w",
+  "984": "f",
+  "985": "p",
+  "986": ".",
+  "987": "b",
+  "988": ",",
+  "989": "v",
+  "990": "k",
+  "991": "'",
+  "992": "I",
+  "993": "T",
+  "994": "A",
+  "995": "S",
+  "996": "x",
+  "997": "W",
+  "998": "j",
+  "999": "B",
+  "1000": "C",
+  "1001": "H",
+  "1002": "?",
+  "1003": "M",
+  "1004": "O",
+  "1005": "Y",
+  "1006": "N",
+  "1007": "P",
+  "1008": "E",
+  "1009": "q",
+  "1010": "L",
+  "1011": "D",
+  "1012": "z",
+  "1013": "G",
+  "1014": "F",
+  "1015": "R",
+  "1016": "!",
+  "1017": "J",
+  "1018": "U",
+  "1019": "K",
+  "1020": "V",
+  "1021": "Q",
+  "1022": "Z",
+  "1023": "X"
+}