diff --git a/src/twn_audio.c b/src/twn_audio.c
index 53d236a..61b20f1 100644
--- a/src/twn_audio.c
+++ b/src/twn_audio.c
@@ -247,6 +247,35 @@ void audio_play(const char *path,
                 float volume,
                 float panning)
 {
+    if (!ctx.audio_initialized) {
+        profile_start("audio initialization");
+
+        SDL_AudioSpec request, got;
+        SDL_zero(request);
+
+        request.freq = AUDIO_FREQUENCY;
+        request.format = AUDIO_F32;
+        request.channels = 2;
+        #ifndef TWN_FEATURE_PUSH_AUDIO
+        request.callback = audio_callback;
+        #endif
+        /* TODO: check for errors */
+        ctx.audio_device = SDL_OpenAudioDevice(NULL, 0, &request, &got, 0);
+        ctx.audio_stream_format = got.format;
+        ctx.audio_stream_frequency = got.freq;
+        ctx.audio_stream_channel_count = got.channels;
+        /* TODO: relax this */
+        SDL_assert_always(got.freq == AUDIO_FREQUENCY);
+        SDL_assert_always(got.format == AUDIO_F32);
+        SDL_assert_always(got.channels == 2);
+
+        SDL_PauseAudioDevice(ctx.audio_device, 0);
+
+        profile_end("audio initialization");
+
+        ctx.audio_initialized = true;
+    }
+
     if (channel) {
         AudioChannelItem *pair = shgetp_null(ctx.audio_channels, channel);
 
diff --git a/src/twn_engine_context_c.h b/src/twn_engine_context_c.h
index ddb9003..35d77db 100644
--- a/src/twn_engine_context_c.h
+++ b/src/twn_engine_context_c.h
@@ -79,6 +79,7 @@ typedef struct EngineContext {
     bool render_double_buffered;
     /* signals mouse focus, used to disable mouse capture */
     bool window_mouse_resident;
+    bool audio_initialized;
 } EngineContext;
 
 /* TODO: does it need to be marked with TWN_API? */
diff --git a/src/twn_loop.c b/src/twn_loop.c
index aea787e..fbadaa9 100644
--- a/src/twn_loop.c
+++ b/src/twn_loop.c
@@ -218,9 +218,11 @@ static void main_loop(void) {
         /* TODO: make it works when ctx.ticks_per_second != 60 */
         #ifdef TWN_FEATURE_PUSH_AUDIO
         static uint8_t audio_buffer[(AUDIO_FREQUENCY / 60) * sizeof (float) * 2];
-        audio_callback(NULL, audio_buffer, sizeof audio_buffer);
-        if (SDL_QueueAudio(ctx.audio_device, audio_buffer, sizeof audio_buffer))
-            CRY_SDL("Error queueing audio: ");
+        if (ctx.audio_initialized) {
+            audio_callback(NULL, audio_buffer, sizeof audio_buffer);
+            if (SDL_QueueAudio(ctx.audio_device, audio_buffer, sizeof audio_buffer))
+                CRY_SDL("Error queueing audio: ");
+        }
         #endif
 
         /* TODO: ctx.game_copy = ctx.game should be placed after it, but it messes with state used in render() */
@@ -537,38 +539,6 @@ static bool initialize(void) {
     /* add a watcher for immediate updates on window size */
     SDL_AddEventWatch(event_callback, NULL);
 
-    /* audio initialization */
-    {
-
-    /* TODO: try to delay it until actually used? */
-    /*       quite a lot of programs might start with silence of resource loading of initial frame */
-    profile_start("audio initialization");
-
-    SDL_AudioSpec request, got;
-    SDL_zero(request);
-
-    request.freq = AUDIO_FREQUENCY;
-    request.format = AUDIO_F32;
-    request.channels = 2;
-    #ifndef TWN_FEATURE_PUSH_AUDIO
-    request.callback = audio_callback;
-    #endif
-    /* TODO: check for errors */
-    ctx.audio_device = SDL_OpenAudioDevice(NULL, 0, &request, &got, 0);
-    ctx.audio_stream_format = got.format;
-    ctx.audio_stream_frequency = got.freq;
-    ctx.audio_stream_channel_count = got.channels;
-    /* TODO: relax this */
-    SDL_assert_always(got.freq == AUDIO_FREQUENCY);
-    SDL_assert_always(got.format == AUDIO_F32);
-    SDL_assert_always(got.channels == 2);
-
-    SDL_PauseAudioDevice(ctx.audio_device, 0);
-
-    profile_end("audio initialization");
-
-    }
-
     /* random seeding */
     /* SDL_GetPerformanceCounter returns some platform-dependent number. */
     /* it should vary between game instances. i checked! random enough for me. */