whisper : default-initialize whisper_mel to avoid uninitialized read (#3981)

whisper_full()/whisper_full_with_state() only compute the mel spectrogram when
n_samples > 0. For n_samples == 0 on a freshly allocated state the mel is never
touched, but struct whisper_mel had no member initializers, so n_len / n_len_org
/ n_mel were indeterminate heap garbage (the state is allocated with
new whisper_state). seek_end is derived from that garbage and, depending on it,
the call either quietly returns 0 or runs the encoder with garbage dimensions
over an empty (NULL) mel buffer, dereferencing address 0 in the mel copy loop.

Give whisper_mel default member initializers so a never-computed mel reads as
0 frames and the n_samples == 0 case deterministically takes the existing
too-short path.

Fixes #3978

src/whisper.cpp

@@ -416,9 +416,9 @@ 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;
+    int n_len     = 0;
+    int n_len_org = 0;
+    int n_mel     = 0;
 
     std::vector<float> data;
 };

tests/CMakeLists.txt

@@ -103,6 +103,16 @@ target_link_libraries(${BUFFER_LOADER_TEST} PRIVATE common)
 add_test(NAME ${BUFFER_LOADER_TEST} COMMAND ${BUFFER_LOADER_TEST})
 set_tests_properties(${BUFFER_LOADER_TEST} PROPERTIES LABELS "unit;gh")
 
+# 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 +1,30 @@
+#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;
+
+    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;
+}