Merge 30e5495e98 into 1fe009caed

src/whisper.cpp

@@ -412,9 +412,12 @@ static const std::map<whisper_alignment_heads_preset, whisper_aheads> g_aheads {
 static std::vector<uint32_t> get_alignment_heads_by_layer(const whisper_context_params & cparams, int il, int32_t n_text_layer, int32_t n_head);
 
 struct whisper_mel {
-    int n_len;
-    int n_len_org;
-    int n_mel;
+    // Default-initialized so a freshly allocated whisper_state whose mel was never
+    // computed (e.g. whisper_full called with n_samples == 0) reads as "0 frames"
+    // instead of indeterminate garbage that can drive a NULL read in the encoder.
+    int n_len     = 0;
+    int n_len_org = 0;
+    int n_mel     = 0;
 
     std::vector<float> data;
 };

tests/CMakeLists.txt

@@ -96,6 +96,16 @@ target_link_libraries(${UTF8_TEST} PRIVATE common)
 add_test(NAME ${UTF8_TEST} COMMAND ${UTF8_TEST})
 set_tests_properties(${UTF8_TEST} PROPERTIES LABELS "unit")
 
+# whisper_full() with n_samples == 0 must not read an uninitialized mel (#3978)
+set(ZERO_SAMPLES_TEST test-whisper-zero-samples)
+add_executable(${ZERO_SAMPLES_TEST} ${ZERO_SAMPLES_TEST}.cpp)
+target_include_directories(${ZERO_SAMPLES_TEST} PRIVATE ../include ../ggml/include ../examples)
+target_link_libraries(${ZERO_SAMPLES_TEST} PRIVATE common)
+target_compile_definitions(${ZERO_SAMPLES_TEST} PRIVATE
+    WHISPER_MODEL_PATH="${PROJECT_SOURCE_DIR}/models/for-tests-ggml-tiny.bin")
+add_test(NAME ${ZERO_SAMPLES_TEST} COMMAND ${ZERO_SAMPLES_TEST})
+set_tests_properties(${ZERO_SAMPLES_TEST} PROPERTIES LABELS "tiny;gh")
+
 # VAD test tests VAD in isolation
 set(VAD_TEST test-vad)
 add_executable(${VAD_TEST} ${VAD_TEST}.cpp)

tests/test-whisper-zero-samples.cpp

@@ -0,0 +1,38 @@
+// Regression test for issue #3978:
+// whisper_full() called with n_samples == 0 on a fresh state must not read the
+// never-computed (previously uninitialized) whisper_mel fields. With the mel
+// default-initialized to "0 frames", the call takes the too-short path and
+// returns cleanly with zero segments instead of running the encoder on garbage
+// dimensions (which could dereference a NULL mel buffer).
+
+#include "whisper.h"
+
+#include <cstdio>
+
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+#include <cassert>
+
+int main() {
+    struct whisper_context_params cparams = whisper_context_default_params();
+    cparams.use_gpu = false;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(WHISPER_MODEL_PATH, cparams);
+    assert(ctx != nullptr);
+
+    struct whisper_full_params params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    params.no_timestamps = true;
+    params.print_progress = false;
+    params.print_realtime = false;
+
+    // n_samples == 0 with a fresh state: the mel is never computed.
+    const int rc = whisper_full(ctx, params, nullptr, 0);
+    assert(rc == 0);
+    assert(whisper_full_n_segments(ctx) == 0);
+
+    whisper_free(ctx);
+
+    printf("test-whisper-zero-samples: OK\n");
+    return 0;
+}