Add a persistent Parakeet helper for low-latency host integrations (#18861)

Factor the Parakeet transcription core into a shared ParakeetTranscriber class and add a persistent parakeet_helper binary (stdin/stdout protocol) for long-lived host integrations. Built in the existing Parakeet CMake presets; README documents the helper workflow.

Merged main and resolved the main.cpp conflict against the TransducerRunner refactor by delegating to ParakeetTranscriber (Token is the shared asr::Token alias). Parakeet CI (test-mlx-parakeet, test-parakeet-xnnpack-linux) passing.

Author: seyeong-han

examples/models/parakeet/CMakeLists.txt

@@ -109,32 +109,49 @@ if(EXECUTORCH_BUILD_VULKAN)
   executorch_target_link_options_shared_lib(vulkan_backend)
 endif()
 
-add_executable(parakeet_runner main.cpp timestamp_utils.cpp tokenizer_utils.cpp)
-if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
-  target_link_options_gc_sections(parakeet_runner)
-  if(NOT APPLE AND NOT MSVC)
-    target_link_options(parakeet_runner PRIVATE "LINKER:-s")
-  endif()
-endif()
+set(parakeet_shared_sources parakeet_transcriber.cpp timestamp_utils.cpp
+                            tokenizer_utils.cpp
+)
 
-# Copy MLX metallib for runtime if MLX delegate is enabled
-if(TARGET mlxdelegate)
-  executorch_target_copy_mlx_metallib(parakeet_runner)
-endif()
+set(parakeet_common_include_directories
+    ${_common_include_directories} ${EXECUTORCH_ROOT}/third-party/json/include
+)
 
-target_include_directories(
-  parakeet_runner PUBLIC ${_common_include_directories}
+add_executable(parakeet_runner main.cpp ${parakeet_shared_sources})
+add_executable(
+  parakeet_helper parakeet_helper.cpp parakeet_helper_protocol.cpp
+                  ${parakeet_shared_sources}
 )
-target_link_libraries(parakeet_runner PUBLIC ${link_libraries})
-target_compile_options(parakeet_runner PUBLIC ${_common_compile_options})
+
+foreach(parakeet_target parakeet_runner parakeet_helper)
+  if(NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
+    target_link_options_gc_sections(${parakeet_target})
+    if(NOT APPLE AND NOT MSVC)
+      target_link_options(${parakeet_target} PRIVATE "LINKER:-s")
+    endif()
+  endif()
+
+  if(TARGET mlxdelegate)
+    executorch_target_copy_mlx_metallib(${parakeet_target})
+  endif()
+
+  target_include_directories(
+    ${parakeet_target} PUBLIC ${parakeet_common_include_directories}
+  )
+  target_link_libraries(${parakeet_target} PUBLIC ${link_libraries})
+  target_compile_options(${parakeet_target} PUBLIC ${_common_compile_options})
+endforeach()
 
 # On Windows, copy required DLLs to the executable directory
 if(MSVC AND EXECUTORCH_BUILD_CUDA)
-  add_custom_command(
-    TARGET parakeet_runner
-    POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E copy_if_different $<TARGET_FILE:aoti_cuda_shims>
-            $<TARGET_FILE_DIR:parakeet_runner>
-    COMMENT "Copying aoti_cuda_shims.dll to parakeet_runner directory"
-  )
+  foreach(parakeet_target parakeet_runner parakeet_helper)
+    add_custom_command(
+      TARGET ${parakeet_target}
+      POST_BUILD
+      COMMAND
+        ${CMAKE_COMMAND} -E copy_if_different $<TARGET_FILE:aoti_cuda_shims>
+        $<TARGET_FILE_DIR:${parakeet_target}>
+      COMMENT "Copying aoti_cuda_shims.dll to ${parakeet_target} directory"
+    )
+  endforeach()
 endif()

examples/models/parakeet/CMakePresets.json

@@ -89,42 +89,42 @@
             "displayName": "Build Parakeet runner (CPU)",
             "configurePreset": "parakeet-cpu",
             "configuration": "Release",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         },
         {
             "name": "parakeet-cuda",
             "displayName": "Build Parakeet runner (CUDA)",
             "configurePreset": "parakeet-cuda",
             "configuration": "Release",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         },
         {
             "name": "parakeet-cuda-debug",
             "displayName": "Build Parakeet runner (CUDA, Debug)",
             "configurePreset": "parakeet-cuda-debug",
             "configuration": "Debug",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         },
         {
             "name": "parakeet-metal",
             "displayName": "Build Parakeet runner (Metal)",
             "configurePreset": "parakeet-metal",
             "configuration": "Release",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         },
         {
             "name": "parakeet-mlx",
             "displayName": "Build Parakeet runner (MLX)",
             "configurePreset": "parakeet-mlx",
             "configuration": "Release",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         },
         {
             "name": "parakeet-vulkan",
             "displayName": "Build Parakeet runner (Vulkan)",
             "configurePreset": "parakeet-vulkan",
             "configuration": "Release",
-            "targets": ["parakeet_runner"]
+            "targets": ["parakeet_runner", "parakeet_helper"]
         }
     ],
     "workflowPresets": [

examples/models/parakeet/README.md

@@ -242,6 +242,11 @@ make parakeet-cuda
 make parakeet-mlx
 ```
 
+Each Parakeet build now produces both:
+
+- `parakeet_runner` for one-shot CLI transcription from an audio file
+- `parakeet_helper` for long-lived host integrations that keep the model warm and stream PCM requests over stdin/stdout
+
 On Windows (PowerShell), use CMake workflow presets directly:
 
 ```powershell
@@ -310,6 +315,26 @@ If your generator is single-config, the runner may be at `.\cmake-out\examples\m
 | `--data_path` | Path to data file (.ptd) for delegate data (required for CUDA/CUDA-Windows) |
 | `--timestamps`     | Timestamp output mode: `none\|token\|word\|segment\|all` (default: `segment`) |
 
+### Persistent Helper
+
+The helper binary uses the same Parakeet transcription stack as `parakeet_runner`,
+but keeps the model loaded across multiple requests so host apps can avoid repeated
+startup and model load overhead.
+
+Example:
+
+```bash
+# Metal
+DYLD_LIBRARY_PATH=/usr/lib ./cmake-out/examples/models/parakeet/parakeet_helper \
+  --model_path examples/models/parakeet/parakeet_metal/model.pte \
+  --tokenizer_path examples/models/parakeet/parakeet_metal/tokenizer.model
+```
+
+The helper accepts framed requests over stdin, validates 16 kHz mono float32 PCM
+payloads, and returns status/result messages over stdout. It is intended for app
+integrations such as the macOS `ExecuWhisper` frontend in the separate
+`executorch-examples` repository.
+
 ### Mobile App
 
 Check out a [demo Android app](https://github.com/meta-pytorch/executorch-examples/tree/main/parakeet/android/ParakeetApp) for Parakeet in the separate `executorch-examples` repository.