diff --git a/UnleashedRecomp/CMakeLists.txt b/UnleashedRecomp/CMakeLists.txt
index e45beef9..239c85c4 100644
--- a/UnleashedRecomp/CMakeLists.txt
+++ b/UnleashedRecomp/CMakeLists.txt
@@ -212,7 +212,6 @@ set(SWA_THIRDPARTY_INCLUDES
     "${SWA_THIRDPARTY_ROOT}/imgui"
     "${SWA_THIRDPARTY_ROOT}/libmspack/libmspack/mspack"
     "${SWA_THIRDPARTY_ROOT}/magic_enum/include"
-    "${SWA_THIRDPARTY_ROOT}/miniaudio"
     "${SWA_THIRDPARTY_ROOT}/stb"
     "${SWA_THIRDPARTY_ROOT}/tiny-AES-c"
     "${SWA_THIRDPARTY_ROOT}/TinySHA1"
diff --git a/UnleashedRecomp/apu/driver/sdl2_driver.cpp b/UnleashedRecomp/apu/driver/sdl2_driver.cpp
index 12bca12f..7c8777e7 100644
--- a/UnleashedRecomp/apu/driver/sdl2_driver.cpp
+++ b/UnleashedRecomp/apu/driver/sdl2_driver.cpp
@@ -7,20 +7,8 @@
 static PPCFunc* g_clientCallback{};
 static uint32_t g_clientCallbackParam{}; // pointer in guest memory
 static SDL_AudioDeviceID g_audioDevice{};
-static std::unique_ptr<GuestThreadContext> g_audioCtx;
-static uint32_t* g_audioOutput;
 static bool g_downMixToStereo;
 
-static void AudioCallback(void*, uint8_t* frames, int len)
-{
-    if (g_audioCtx == nullptr)
-        g_audioCtx = std::make_unique<GuestThreadContext>(0);
-
-    g_audioCtx->ppcContext.r3.u64 = g_clientCallbackParam;
-    g_audioOutput = reinterpret_cast<uint32_t*>(frames);
-    (*g_clientCallback)(g_audioCtx->ppcContext, reinterpret_cast<uint8_t*>(g_memory.base));
-}
-
 void XAudioInitializeSystem()
 {
     SDL_SetHint(SDL_HINT_AUDIO_CATEGORY, "playback");
@@ -32,7 +20,6 @@ void XAudioInitializeSystem()
     desired.format = AUDIO_F32SYS;
     desired.channels = XAUDIO_NUM_CHANNELS;
     desired.samples = XAUDIO_NUM_SAMPLES;
-    desired.callback = AudioCallback;
     g_audioDevice = SDL_OpenAudioDevice(nullptr, 0, &desired, &obtained, SDL_AUDIO_ALLOW_CHANNELS_CHANGE);
 
     if (obtained.channels != 2 && obtained.channels != XAUDIO_NUM_CHANNELS)
@@ -44,6 +31,44 @@ void XAudioInitializeSystem()
     g_downMixToStereo = (obtained.channels == 2);
 }
 
+static std::unique_ptr<std::thread> g_audioThread;
+
+static void AudioThread()
+{
+    using namespace std::chrono_literals;
+
+    GuestThreadContext ctx(0);
+
+    size_t channels = g_downMixToStereo ? 2 : XAUDIO_NUM_CHANNELS;
+
+    constexpr double INTERVAL = double(XAUDIO_NUM_SAMPLES) / double(XAUDIO_SAMPLES_HZ);
+    auto start = std::chrono::steady_clock::now();
+    size_t iteration = 1;
+
+    while (true)
+    {
+        uint32_t queuedAudioSize = SDL_GetQueuedAudioSize(g_audioDevice);
+        constexpr size_t MAX_LATENCY = 10;
+        const size_t callbackAudioSize = channels * XAUDIO_NUM_SAMPLES * sizeof(float);
+
+        if ((queuedAudioSize / callbackAudioSize) <= MAX_LATENCY)
+        {
+            ctx.ppcContext.r3.u32 = g_clientCallbackParam;
+            g_clientCallback(ctx.ppcContext, reinterpret_cast<uint8_t*>(g_memory.base));
+        }
+
+        auto next = start + std::chrono::duration<double>(iteration * INTERVAL);
+        auto now = std::chrono::steady_clock::now();
+
+        if ((next - now) > 1s)
+            next = now;
+
+        std::this_thread::sleep_until(next);
+
+        iteration = std::chrono::duration<double>(std::chrono::steady_clock::now() - start).count() / INTERVAL + 1;
+    }
+}
+
 void XAudioRegisterClient(PPCFunc* callback, uint32_t param)
 {
     auto* pClientParam = static_cast<uint32_t*>(g_userHeap.Alloc(sizeof(param)));
@@ -53,6 +78,7 @@ void XAudioRegisterClient(PPCFunc* callback, uint32_t param)
     g_clientCallback = callback;
 
     SDL_PauseAudioDevice(g_audioDevice, 0);
+    g_audioThread = std::make_unique<std::thread>(AudioThread);
 }
 
 void XAudioSubmitFrame(void* samples)
@@ -68,6 +94,8 @@ void XAudioSubmitFrame(void* samples)
 
         auto floatSamples = reinterpret_cast<be<float>*>(samples);
 
+        std::array<float, 2 * XAUDIO_NUM_SAMPLES> audioFrames;
+
         for (size_t i = 0; i < XAUDIO_NUM_SAMPLES; i++)
         {
             float ch0 = floatSamples[0 * XAUDIO_NUM_SAMPLES + i];
@@ -77,21 +105,24 @@ void XAudioSubmitFrame(void* samples)
             float ch4 = floatSamples[4 * XAUDIO_NUM_SAMPLES + i];
             float ch5 = floatSamples[5 * XAUDIO_NUM_SAMPLES + i];
 
-            float left  = ch0 + ch2 * 0.75f + ch4;
-            float right = ch1 + ch2 * 0.75f + ch5;
-
-            g_audioOutput[i * 2 + 0] = *reinterpret_cast<uint32_t*>(&left);
-            g_audioOutput[i * 2 + 1] = *reinterpret_cast<uint32_t*>(&right);
+            audioFrames[i * 2 + 0] = ch0 + ch2 * 0.75f + ch4;
+            audioFrames[i * 2 + 1] = ch1 + ch2 * 0.75f + ch5;
         }
+
+        SDL_QueueAudio(g_audioDevice, &audioFrames, sizeof(audioFrames));
     }
     else
     {
         auto rawSamples = reinterpret_cast<be<uint32_t>*>(samples);
 
+        std::array<uint32_t, XAUDIO_NUM_CHANNELS * XAUDIO_NUM_SAMPLES> audioFrames;
+
         for (size_t i = 0; i < XAUDIO_NUM_SAMPLES; i++)
         {
             for (size_t j = 0; j < XAUDIO_NUM_CHANNELS; j++)
-                g_audioOutput[i * XAUDIO_NUM_CHANNELS + j] = rawSamples[j * XAUDIO_NUM_SAMPLES + i];
+                audioFrames[i * XAUDIO_NUM_CHANNELS + j] = rawSamples[j * XAUDIO_NUM_SAMPLES + i];
         }
+
+        SDL_QueueAudio(g_audioDevice, &audioFrames, sizeof(audioFrames));
     }
 }
diff --git a/UnleashedRecomp/cpu/guest_thread.cpp b/UnleashedRecomp/cpu/guest_thread.cpp
index 32412b40..290fc51c 100644
--- a/UnleashedRecomp/cpu/guest_thread.cpp
+++ b/UnleashedRecomp/cpu/guest_thread.cpp
@@ -32,6 +32,7 @@ GuestThreadContext::GuestThreadContext(uint32_t cpuNumber)
     ppcContext.fn = (uint8_t*)g_codeCache.bucket;
     ppcContext.r1.u64 = g_memory.MapVirtual(thread + PCR_SIZE + TLS_SIZE + TEB_SIZE + STACK_SIZE); // stack pointer
     ppcContext.r13.u64 = g_memory.MapVirtual(thread);
+    ppcContext.fpscr.loadFromHost();
 
     assert(GetPPCContext() == nullptr);
     SetPPCContext(ppcContext);
@@ -77,7 +78,7 @@ uint32_t GuestThread::Start(const GuestThreadParams& params)
     GuestThreadContext ctx(cpuNumber);
     ctx.ppcContext.r3.u64 = params.value;
 
-    GuestCode::Run(g_codeCache.Find(params.function), &ctx.ppcContext, g_memory.Translate(0));
+    reinterpret_cast<PPCFunc*>(g_codeCache.Find(params.function))(ctx.ppcContext, reinterpret_cast<uint8_t*>(g_memory.base));
 
     return ctx.ppcContext.r3.u32;
 }
diff --git a/UnleashedRecomp/kernel/imports.cpp b/UnleashedRecomp/kernel/imports.cpp
index d570d16a..e206b47b 100644
--- a/UnleashedRecomp/kernel/imports.cpp
+++ b/UnleashedRecomp/kernel/imports.cpp
@@ -76,16 +76,16 @@ static std::atomic<uint32_t> g_keSetEventGeneration;
 
 struct Semaphore final : KernelObject, HostObject<XKSEMAPHORE>
 {
-    std::atomic<uint32_t> count;
+    std::counting_semaphore<> semaphore;
     uint32_t maximumCount;
 
     Semaphore(XKSEMAPHORE* semaphore)
-        : count(semaphore->Header.SignalState), maximumCount(semaphore->Limit)
+        : semaphore(semaphore->Header.SignalState), maximumCount(semaphore->Limit)
     {
     }
 
     Semaphore(uint32_t count, uint32_t maximumCount)
-        : count(count), maximumCount(maximumCount)
+        : semaphore(count), maximumCount(maximumCount)
     {
     }
 
@@ -93,32 +93,11 @@ struct Semaphore final : KernelObject, HostObject<XKSEMAPHORE>
     {
         if (timeout == 0)
         {
-            uint32_t currentCount = count.load();
-            if (currentCount != 0)
-            {
-                if (count.compare_exchange_weak(currentCount, currentCount - 1))
-                    return STATUS_SUCCESS;
-            }
-
-            return STATUS_TIMEOUT;
+            return semaphore.try_acquire() ? STATUS_SUCCESS : STATUS_TIMEOUT;
         }
         else if (timeout == INFINITE)
         {
-            uint32_t currentCount;
-            while (true)
-            {
-                currentCount = count.load();
-                if (currentCount != 0)
-                {
-                    if (count.compare_exchange_weak(currentCount, currentCount - 1))
-                        return STATUS_SUCCESS;
-                }
-                else
-                {
-                    count.wait(0);
-                }
-            }
-
+            semaphore.acquire();
             return STATUS_SUCCESS;
         }
         else
@@ -130,13 +109,7 @@ struct Semaphore final : KernelObject, HostObject<XKSEMAPHORE>
 
     void Release(uint32_t releaseCount, uint32_t* previousCount)
     {
-        if (previousCount != nullptr)
-            *previousCount = count;
-
-        assert(count + releaseCount <= maximumCount);
-
-        count += releaseCount;
-        count.notify_all();
+        semaphore.release(releaseCount);
     }
 };