speech addons: enable per-arch ggml CPU variants on desktop linux-arm64 (whispercpp, bci, tts-ggml)

packages/bci-whispercpp/CHANGELOG.md

@@ -10,6 +10,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Changed
 
 - Faster desktop neural-signal preprocessing on the GPU (Vulkan) and CPU paths. Public API and transcription output are unchanged.
+- Desktop linux-arm64 prebuilds now ship per-arch ggml CPU variants (`whisper-cpp` override 1.9.1#3, pulling `ggml-speech` 2026-07-14): the previous armv8-a-baseline build compiled out the ARM dotprod/fp16 kernels. The addon now loads the dynamically-loadable ggml backends on linux-arm64 (previously Android-only).
 
 ## [0.5.0] - 2026-07-14
 

packages/bci-whispercpp/addon/src/model-interface/bci/BCIModel.cpp

@@ -32,9 +32,10 @@ static bool shouldAbortWhisper(void* userData) {
          cancelRequested->load(std::memory_order_relaxed);
 }
 
-#if defined(__ANDROID__)
+#if defined(__ANDROID__) || (defined(__linux__) && defined(__aarch64__))
 namespace {
-// Android ships ggml with `GGML_BACKEND_DL=ON`, so no backend is
+// Android -- and, since ggml-speech 2026-07-14, desktop linux-arm64
+// prebuilds too -- ship ggml with `GGML_BACKEND_DL=ON`, so no backend is
 // statically registered. dlopen the per-arch CPU + GPU `.so` modules
 // once per process; otherwise whisper_init aborts on a NULL CPU
 // device. Mirrors transcription-whispercpp / llm-llamacpp /
@@ -50,13 +51,13 @@ namespace {
 // `GGML_BACKEND_DL=ON` flag; from that point on this function is
 // the only thing standing between us and the SIGABRT we already
 // hit on transcription's PR #2124.
-void ensureBackendsLoadedAndroid(const std::string& backendsDir) {
+void ensureBackendsLoaded(const std::string& backendsDir) {
   static std::once_flag flag;
   std::call_once(flag, [&]() {
     if (backendsDir.empty()) {
       QLOG(
           qvac_lib_inference_addon_cpp::logger::Priority::WARNING,
-          "Android: configurationParams.backendsDir not set; falling back to "
+          "configurationParams.backendsDir not set; falling back to "
           "ggml_backend_load_all() (default search path). CPU / Vulkan / "
           "OpenCL registration may fail inside an APK with default "
           "compressed-native-libs packaging.");
@@ -73,13 +74,13 @@ void ensureBackendsLoadedAndroid(const std::string& backendsDir) {
 #endif
     QLOG(
         qvac_lib_inference_addon_cpp::logger::Priority::INFO,
-        std::string("Android: loading ggml backends from: ") +
+        std::string("loading ggml backends from: ") +
             variantsDir.string());
     ggml_backend_load_all_from_path(variantsDir.string().c_str());
   });
 }
 } // namespace
-#endif // __ANDROID__
+#endif // __ANDROID__ || linux-arm64
 
 BCIModel::BCIModel(BCIConfig config)
     : cfg_(std::move(config)), neuralProcessor_() {}
@@ -187,8 +188,8 @@ int adrenoOpenclGpuDeviceIndex() {
 void BCIModel::load() {
   if (ctx_) return;
 
-#if defined(__ANDROID__)
-  ensureBackendsLoadedAndroid(cfg_.backendsDir);
+#if defined(__ANDROID__) || (defined(__linux__) && defined(__aarch64__))
+  ensureBackendsLoaded(cfg_.backendsDir);
 #endif
 
   whisper_context_params contextParams = toWhisperContextParams(cfg_);

packages/bci-whispercpp/vcpkg.json

@@ -38,7 +38,7 @@
     {
       "name": "whisper-cpp",
       "version": "1.9.1",
-      "port-version": 1
+      "port-version": 3
     }
   ]
 }

packages/transcription-whispercpp/CHANGELOG.md

@@ -41,6 +41,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   objects, and split large C++/JS functions (and their inlined loops) into
   smaller named helpers per the team coding standards. These are internal
   refactors with no public API change.
+- Desktop linux-arm64 prebuilds now ship per-arch ggml CPU variants (`whisper-cpp` override 1.9.1#3, pulling `ggml-speech` 2026-07-14): the previous armv8-a-baseline build compiled out the ARM dotprod/fp16 kernels (base f16 mean RTF 0.332 -> 0.097, base q8_0 0.127 -> 0.063, small f16 1.345 -> 0.343 on ubuntu-24.04-arm). The addon now loads the dynamically-loadable ggml backends on linux-arm64 (previously Android-only).
 
 ## [0.12.0] - 2026-07-14
 

packages/transcription-whispercpp/addon/src/model-interface/whisper.cpp/WhisperModel.cpp

@@ -73,19 +73,20 @@ auto WhisperModel::formatCaptionOutput(Transcript& transcript) -> void {
                     std::to_string(static_cast<int>(transcript.end)) + "|>";
 }
 
-#if defined(__ANDROID__)
+#if defined(__ANDROID__) || (defined(__linux__) && defined(__aarch64__))
 namespace {
-// Android ships ggml with `GGML_BACKEND_DL=ON`, so no backend is
+// Android -- and, since ggml-speech 2026-07-14, desktop linux-arm64
+// prebuilds too -- ship ggml with `GGML_BACKEND_DL=ON`, so no backend is
 // statically registered. dlopen the per-arch CPU + GPU `.so` modules
 // once per process; otherwise whisper_init aborts on a NULL CPU device.
 // Mirrors packages/{diffusion-cpp,llm-llamacpp,classification-ggml,…}.
-void ensureBackendsLoadedAndroid(const std::string& backendsDir) {
+void ensureBackendsLoaded(const std::string& backendsDir) {
   static std::once_flag flag;
   std::call_once(flag, [&]() {
     if (backendsDir.empty()) {
       QLOG(
           qvac_lib_inference_addon_cpp::logger::Priority::WARNING,
-          "Android: configurationParams.backendsDir not set; falling back to "
+          "configurationParams.backendsDir not set; falling back to "
           "ggml_backend_load_all() (default search path). CPU/Vulkan/OpenCL "
           "registration may fail inside an APK.");
       ggml_backend_load_all();
@@ -101,13 +102,13 @@ void ensureBackendsLoadedAndroid(const std::string& backendsDir) {
 #endif
     QLOG(
         qvac_lib_inference_addon_cpp::logger::Priority::INFO,
-        std::string("Android: loading ggml backends from: ") +
+        std::string("loading ggml backends from: ") +
             variantsDir.string());
     ggml_backend_load_all_from_path(variantsDir.string().c_str());
   });
 }
 } // namespace
-#endif // __ANDROID__
+#endif // __ANDROID__ || linux-arm64
 
 namespace {
 std::string toLowerCopy(std::string value) {
@@ -202,8 +203,8 @@ int adrenoOpenclGpuDeviceIndex() {
 void WhisperModel::load() {
   if (!ctx_) {
 
-#if defined(__ANDROID__)
-    ensureBackendsLoadedAndroid(cfg_.backendsDir);
+#if defined(__ANDROID__) || (defined(__linux__) && defined(__aarch64__))
+    ensureBackendsLoaded(cfg_.backendsDir);
 #endif
 
     whisper_context_params contextParams = toWhisperContextParams(cfg_);

packages/transcription-whispercpp/vcpkg.json

@@ -29,7 +29,7 @@
     {
       "name": "whisper-cpp",
       "version": "1.9.1",
-      "port-version": 1
+      "port-version": 3
     }
   ]
 }

packages/tts-ggml/CHANGELOG.md

@@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- Desktop linux-arm64 prebuilds now ship per-arch ggml CPU variants (`tts-cpp` >= 2026-07-13#2, pulling `ggml-speech` 2026-07-14): the previous armv8-a-baseline build compiled out the ARM dotprod/fp16 kernels (chatterbox q4 CPU mean RTF 2.70 -> 2.28 on ubuntu-24.04-arm).
+
 ### Added
 
 - **Chatterbox S3Gen classifier-free-guidance rate (`cfgRate`) (QVAC-21908).**

packages/tts-ggml/vcpkg.json

@@ -10,7 +10,7 @@
     },
     {
       "name": "tts-cpp",
-      "version>=": "2026-07-13#1"
+      "version>=": "2026-07-13#2"
     }
   ],
   "features": {