build: allow embedding voice-detect.cpp in another ggml project

CMakeLists.txt

@@ -16,6 +16,16 @@ option(VOICEDETECT_GGML_METAL  "Forward GGML_METAL" OFF)
 option(VOICEDETECT_GGML_VULKAN "Forward GGML_VULKAN" OFF)
 option(VOICEDETECT_GGML_HIP    "Forward GGML_HIP (ROCm)" OFF)
 
+set(VOICEDETECT_TOP_LEVEL OFF)
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+  set(VOICEDETECT_TOP_LEVEL ON)
+endif()
+option(VOICEDETECT_EXTERNAL_DR_WAV "Use the embedding project's dr_wav implementation" OFF)
+
+# An embedding host owns its ggml configuration; everything up to the ggml
+# add_subdirectory below only applies to a top-level build.
+if(VOICEDETECT_TOP_LEVEL)
+
 # Multi-ISA CPU backend with RUNTIME dispatch. ggml can build one CPU backend
 # per x86 micro-arch (sse42/sandybridge/haswell/skylakex/icelake/zen4/...) as
 # dlopen'd modules and pick the best one the host actually supports AT LOAD TIME.
@@ -130,7 +140,11 @@ if(BASH_EXECUTABLE AND EXISTS "${CMAKE_SOURCE_DIR}/scripts/apply_ggml_patches.sh
   endif()
 endif()
 
-add_subdirectory(third_party/ggml)
+endif() # VOICEDETECT_TOP_LEVEL
+
+if(NOT TARGET ggml)
+  add_subdirectory(third_party/ggml)
+endif()
 
 set(VOICEDETECT_SRC
     src/voicedetect.cpp
@@ -159,9 +173,12 @@ if(VOICEDETECT_SHARED)
 else()
   add_library(voicedetect STATIC ${VOICEDETECT_SRC})
 endif()
-target_include_directories(voicedetect PUBLIC include PRIVATE src ${CMAKE_SOURCE_DIR}/third_party)
+target_include_directories(voicedetect PUBLIC include PRIVATE src ${CMAKE_CURRENT_SOURCE_DIR}/third_party)
 target_compile_definitions(voicedetect PUBLIC $<$<CXX_COMPILER_ID:MSVC>:_USE_MATH_DEFINES>)
 target_link_libraries(voicedetect PUBLIC ggml)
+if(VOICEDETECT_EXTERNAL_DR_WAV)
+  target_compile_definitions(voicedetect PRIVATE VOICEDETECT_EXTERNAL_DR_WAV)
+endif()
 if(VOICEDETECT_GGML_CUDNN)
   target_compile_definitions(voicedetect PRIVATE VOICEDETECT_GGML_CUDNN)
 endif()
@@ -218,10 +235,10 @@ if(VOICEDETECT_HAS_AVX2)
   endif()
 endif()
 
-if(VOICEDETECT_BUILD_CLI)
+if(VOICEDETECT_BUILD_CLI AND VOICEDETECT_TOP_LEVEL)
   add_subdirectory(examples/cli)
 endif()
-if(VOICEDETECT_BUILD_TESTS)
+if(VOICEDETECT_BUILD_TESTS AND VOICEDETECT_TOP_LEVEL)
   enable_testing()
   add_subdirectory(tests)
 endif()

examples/cli/CMakeLists.txt

@@ -1,3 +1,3 @@
 add_executable(voicedetect-cli main.cpp)
 target_link_libraries(voicedetect-cli PRIVATE voicedetect)
-target_include_directories(voicedetect-cli PRIVATE ${CMAKE_SOURCE_DIR}/src)
+target_include_directories(voicedetect-cli PRIVATE ${PROJECT_SOURCE_DIR}/src)

include/voicedetect_capi.h

@@ -88,6 +88,11 @@ int voicedetect_capi_verify_paths(voicedetect_ctx* ctx, const char* a,
 char* voicedetect_capi_analyze_path_json(voicedetect_ctx* ctx,
                                          const char* wav_path);
 
+// Size of the L2-normalized embedding this model produces (192 to 512), 0 for a
+// model with no speaker embedding (the age/gender/emotion analyze heads), -1 for
+// a NULL ctx. Additive: no ABI version bump.
+int voicedetect_capi_embedding_dim(voicedetect_ctx* ctx);
+
 #ifdef __cplusplus
 } // extern "C"
 #endif

src/audio_io.cpp

@@ -1,4 +1,6 @@
-#define DR_WAV_IMPLEMENTATION
+#ifndef VOICEDETECT_EXTERNAL_DR_WAV
+#define DR_WAV_IMPLEMENTATION  // an embedding project may provide it instead
+#endif
 #include "dr_wav.h"
 #include "audio_io.hpp"
 #include "common.hpp"

src/voicedetect_capi.cpp

@@ -62,6 +62,11 @@ extern "C" int voicedetect_capi_abi_version(void) {
     return VOICEDETECT_CAPI_ABI_VERSION;
 }
 
+extern "C" int voicedetect_capi_embedding_dim(voicedetect_ctx* ctx) {
+    if (!ctx || !ctx->model) return -1;
+    return ctx->model->embedding_dim();
+}
+
 extern "C" voicedetect_ctx* voicedetect_capi_load(const char* gguf_path) {
     if (!gguf_path) return nullptr;
     try {

tests/CMakeLists.txt

@@ -1,12 +1,13 @@
 function(vd_add_test name)
   add_executable(${name} ${name}.cpp)
   target_link_libraries(${name} PRIVATE voicedetect)
-  target_include_directories(${name} PRIVATE ${CMAKE_SOURCE_DIR}/src ${CMAKE_SOURCE_DIR}/third_party)
+  target_include_directories(${name} PRIVATE ${PROJECT_SOURCE_DIR}/src ${PROJECT_SOURCE_DIR}/third_party)
   add_test(NAME ${name} COMMAND ${name})
   set_tests_properties(${name} PROPERTIES SKIP_RETURN_CODE 77)
 endfunction()
 
 vd_add_test(test_smoke)
+vd_add_test(test_capi_dim)
 vd_add_test(test_fbank)
 vd_add_test(test_encoder)
 vd_add_test(test_embed)
@@ -19,37 +20,37 @@ vd_add_test(test_cache_safety)
 # Manual benchmark target (not registered as a ctest; run directly).
 add_executable(bench_wespeaker bench_wespeaker.cpp)
 target_link_libraries(bench_wespeaker PRIVATE voicedetect)
-target_include_directories(bench_wespeaker PRIVATE ${CMAKE_SOURCE_DIR}/src ${CMAKE_SOURCE_DIR}/third_party)
+target_include_directories(bench_wespeaker PRIVATE ${PROJECT_SOURCE_DIR}/src ${PROJECT_SOURCE_DIR}/third_party)
 
 add_executable(bench_eres2net bench_eres2net.cpp)
 target_link_libraries(bench_eres2net PRIVATE voicedetect)
-target_include_directories(bench_eres2net PRIVATE ${CMAKE_SOURCE_DIR}/src ${CMAKE_SOURCE_DIR}/third_party)
+target_include_directories(bench_eres2net PRIVATE ${PROJECT_SOURCE_DIR}/src ${PROJECT_SOURCE_DIR}/third_party)
 
 # Model/baseline-dependent tests read fixtures/baselines via paths relative to
 # the project root, and skip (exit 77) when the baseline env var is unset.
 set_tests_properties(test_fbank test_encoder test_embed test_verify test_analyze test_age_gender test_cache_safety PROPERTIES
                      LABELS "model"
-                     WORKING_DIRECTORY ${CMAKE_SOURCE_DIR})
+                     WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
 
 # Python converter check (skips with exit 77 when the venv/model are absent).
-find_program(VOICEDETECT_PY NAMES python3 HINTS ${CMAKE_SOURCE_DIR}/.venv/bin)
+find_program(VOICEDETECT_PY NAMES python3 HINTS ${PROJECT_SOURCE_DIR}/.venv/bin)
 if(VOICEDETECT_PY)
-  add_test(NAME check_convert COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_convert.py)
+  add_test(NAME check_convert COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_convert.py)
   set_tests_properties(check_convert PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
   # Analyze (wav2vec2 emotion) converter check: a SEPARATE analyze GGUF, skips
   # independently of the speaker-encoder check when its HF model is absent.
-  add_test(NAME check_convert_analyze COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_convert.py --analyze)
+  add_test(NAME check_convert_analyze COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_convert.py --analyze)
   set_tests_properties(check_convert_analyze PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
-  add_test(NAME check_baseline_analyze COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_baseline.py --analyze)
+  add_test(NAME check_baseline_analyze COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_baseline.py --analyze)
   set_tests_properties(check_baseline_analyze PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
   # Age/gender (audeering wav2vec2-large-robust) converter + baseline checks: a
   # SEPARATE analyze GGUF, skips independently when its HF model is absent.
-  add_test(NAME check_convert_age_gender COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_convert.py --age-gender)
+  add_test(NAME check_convert_age_gender COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_convert.py --age-gender)
   set_tests_properties(check_convert_age_gender PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
-  add_test(NAME check_baseline_age_gender COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_baseline.py --age-gender)
+  add_test(NAME check_baseline_age_gender COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_baseline.py --age-gender)
   set_tests_properties(check_baseline_age_gender PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
-  add_test(NAME check_manifest COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_manifest.py)
+  add_test(NAME check_manifest COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_manifest.py)
   set_tests_properties(check_manifest PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
-  add_test(NAME check_baseline COMMAND ${VOICEDETECT_PY} ${CMAKE_SOURCE_DIR}/tests/python/check_baseline.py)
+  add_test(NAME check_baseline COMMAND ${VOICEDETECT_PY} ${PROJECT_SOURCE_DIR}/tests/python/check_baseline.py)
   set_tests_properties(check_baseline PROPERTIES SKIP_RETURN_CODE 77 LABELS "model")
 endif()

tests/test_capi_dim.cpp

@@ -0,0 +1,32 @@
+// voicedetect_capi_embedding_dim: NULL ctx is -1; with VOICEDETECT_TEST_GGUF set,
+// it matches the length embed_pcm returns. Skips (77) without a model.
+#include "voicedetect_capi.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+int main() {
+    if (voicedetect_capi_embedding_dim(nullptr) != -1) {
+        std::fprintf(stderr, "FAIL: NULL ctx should give -1\n");
+        return 1;
+    }
+    const char* path = std::getenv("VOICEDETECT_TEST_GGUF");
+    if (!path) return 77;
+    voicedetect_ctx* ctx = voicedetect_capi_load(path);
+    if (!ctx) { std::fprintf(stderr, "FAIL: load %s\n", path); return 1; }
+    const int dim = voicedetect_capi_embedding_dim(ctx);
+    std::vector<float> pcm(32000);
+    for (size_t i = 0; i < pcm.size(); ++i) pcm[i] = 0.1f * (float)((i * 37) % 200 - 100) / 100.0f;
+    float* v = nullptr;
+    int n = 0;
+    const int rc = voicedetect_capi_embed_pcm(ctx, pcm.data(), (int)pcm.size(), 16000, &v, &n);
+    voicedetect_capi_free_vec(v);
+    voicedetect_capi_free(ctx);
+    if (rc != 0 || dim <= 0 || dim != n) {
+        std::fprintf(stderr, "FAIL: rc=%d dim=%d embed len=%d\n", rc, dim, n);
+        return 1;
+    }
+    std::printf("embedding_dim OK (%d)\n", dim);
+    return 0;
+}