Merge branch 'master' of https://github.com/ggerganov/whisper.cpp into miniaudio

Author: data-man

.github/workflows/docker.yml

@@ -28,6 +28,8 @@ jobs:
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v3
+        with:
+          image: tonistiigi/binfmt:qemu-v7.0.0-28
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3

examples/stream/stream.cpp

@@ -23,6 +23,7 @@ struct whisper_params {
     int32_t capture_id = -1;
     int32_t max_tokens = 32;
     int32_t audio_ctx  = 0;
+    int32_t beam_size  = -1;
 
     float vad_thold    = 0.6f;
     float freq_thold   = 100.0f;
@@ -59,6 +60,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         else if (arg == "-c"    || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
         else if (arg == "-mt"   || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
         else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")     { params.beam_size     = std::stoi(argv[++i]); }
         else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
         else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
         else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
@@ -96,6 +98,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                              params.capture_id);
     fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",       params.max_tokens);
     fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -bs N,    --beam-size N   [%-7d] beam size for beam search\n",                      params.beam_size);
     fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
     fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
     fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
@@ -298,7 +301,7 @@ int main(int argc, char ** argv) {
 
         // run the inference
         {
-            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+            whisper_full_params wparams = whisper_full_default_params(params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY);
 
             wparams.print_progress   = false;
             wparams.print_special    = params.print_special;
@@ -309,6 +312,7 @@ int main(int argc, char ** argv) {
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();
             wparams.n_threads        = params.n_threads;
+            wparams.beam_search.beam_size = params.beam_size;
 
             wparams.audio_ctx        = params.audio_ctx;
 

src/CMakeLists.txt

@@ -94,6 +94,10 @@ set_target_properties(whisper PROPERTIES
 target_include_directories(whisper PUBLIC . ../include)
 target_compile_features   (whisper PUBLIC cxx_std_11) # don't bump
 
+if (CMAKE_CXX_BYTE_ORDER STREQUAL "BIG_ENDIAN")
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_BIG_ENDIAN)
+endif()
+
 if (WHISPER_EXTRA_FLAGS)
     target_compile_options(whisper PRIVATE ${WHISPER_EXTRA_FLAGS})
 endif()

src/whisper.cpp

@@ -39,17 +39,17 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if defined(GGML_BIG_ENDIAN)
-#include <bit>
-
+#if defined(WHISPER_BIG_ENDIAN)
 template<typename T>
 static T byteswap(T value) {
-    return std::byteswap(value);
-}
-
-template<>
-float byteswap(float value) {
-    return std::bit_cast<float>(byteswap(std::bit_cast<std::uint32_t>(value)));
+    T value_swapped;
+    char * source = reinterpret_cast<char *>(&value);
+    char * target = reinterpret_cast<char *>(&value_swapped);
+    int size = sizeof(T);
+    for (int i = 0; i < size; i++) {
+        target[size - 1 - i] = source[i];
+    }
+    return value_swapped;
 }
 
 template<typename T>
@@ -85,14 +85,14 @@ static void byteswap_tensor(ggml_tensor * tensor) {
 }
 
 #define BYTESWAP_VALUE(d) d = byteswap(d)
-#define BYTESWAP_FILTERS(f)            \
+#define BYTESWAP_FILTERS(f)           \
     do {                              \
         for (auto & datum : f.data) { \
             datum = byteswap(datum);  \
         }                             \
     } while (0)
-#define BYTESWAP_TENSOR(t)       \
-    do {                         \
+#define BYTESWAP_TENSOR(t)  \
+    do {                    \
         byteswap_tensor(t); \
     } while (0)
 #else