diff --git a/SConstruct b/SConstruct
index d42a40b7..cc902a59 100644
--- a/SConstruct
+++ b/SConstruct
@@ -30,7 +30,7 @@ print "platform: ", env['PLATFORM']
 if env['PLATFORM'] == 'cygwin':
   env.Append(CCFLAGS = " -mno-cygwin")
   env.Append(LINKFLAGS = " -mno-cygwin")
-  env['LIBS'] = ['ddraw','dinput','dsound','gdi32','dxguid','winmm','shell32','wsock32','comdlg32','ole32'];
+  env['LIBS'] = ['wsock32'];
 
 conf = Configure(env)
 if not conf.CheckLib('SDL'):
@@ -60,6 +60,8 @@ if env['FRAMESKIP']:
 # parse SDL cflags/libs
 env.ParseConfig('sdl-config --cflags --libs')
 
+print "LIBS:",env['CCFLAGS']
+#env['IBS'] = ['wsock32', 'SDL', 'z', 'mingw32', 'SDL']
 Export('env')
 
 SConscript(['src/SConscript'])
diff --git a/TODO-PROJECT b/TODO-PROJECT
index 394918b3..f2d53513 100644
--- a/TODO-PROJECT
+++ b/TODO-PROJECT
@@ -2,6 +2,8 @@ Items to be completed before 2.0 release
 
 FASTAPASS / FP_FASTAPASS / Are these archaic? They suck - ??
 
+Separate frameskip/pause from EmulationPaused - zeromus
+
 Make ALL Debugging code conditional - zeromus
 
 Doxygen integration - DONE
diff --git a/src/drivers/common/Makefile.am.inc b/src/drivers/common/Makefile.am.inc
deleted file mode 100644
index 626a617d..00000000
--- a/src/drivers/common/Makefile.am.inc
+++ /dev/null
@@ -1 +0,0 @@
-fceu_SOURCES	+=	drivers/common/args.cpp drivers/common/cheat.cpp drivers/common/config.cpp drivers/common/vidblit.cpp drivers/common/hq2x.cpp drivers/common/hq3x.cpp drivers/common/scale2x.cpp drivers/common/scale3x.cpp drivers/common/scalebit.cpp
diff --git a/src/drivers/win/OutputDS.cpp b/src/drivers/win/OutputDS.cpp
new file mode 100644
index 00000000..b0139518
--- /dev/null
+++ b/src/drivers/win/OutputDS.cpp
@@ -0,0 +1,288 @@
+#include "oakra.h"
+
+#include "dsound.h"
+#include <vector>
+
+#define LATENCY_MS (50)
+
+
+int voltbl[] = {
+-10000,-8750,-8018,-7499,-7097,-6768,-6490,-6250,-6037,-5847,-5675,-5518,-5374,-5240,-5116,-5000,
+-4890,-4787,-4690,-4597,-4509,-4425,-4345,-4268,-4195,-4124,-4056,-3990,-3927,-3866,-3807,-3749,
+-3694,-3640,-3588,-3537,-3488,-3440,-3393,-3347,-3303,-3259,-3217,-3175,-3135,-3095,-3056,-3018,
+-2981,-2945,-2909,-2874,-2840,-2806,-2773,-2740,-2708,-2677,-2646,-2616,-2586,-2557,-2528,-2500,
+-2472,-2444,-2417,-2390,-2364,-2338,-2312,-2287,-2262,-2238,-2213,-2190,-2166,-2143,-2120,-2097,
+-2075,-2053,-2031,-2009,-1988,-1967,-1946,-1925,-1905,-1885,-1865,-1845,-1826,-1806,-1787,-1768,
+-1750,-1731,-1713,-1695,-1677,-1659,-1641,-1624,-1607,-1590,-1573,-1556,-1539,-1523,-1506,-1490,
+-1474,-1458,-1443,-1427,-1412,-1396,-1381,-1366,-1351,-1336,-1321,-1307,-1292,-1278,-1264,-1250,
+-1235,-1222,-1208,-1194,-1180,-1167,-1153,-1140,-1127,-1114,-1101,-1088,-1075,-1062,-1050,-1037,
+-1025,-1012,-1000,-988,-976,-963,-951,-940,-928,-916,-904,-893,-881,-870,-858,-847,
+-836,-825,-814,-803,-792,-781,-770,-759,-748,-738,-727,-717,-706,-696,-685,-675,
+-665,-655,-645,-635,-625,-615,-605,-595,-585,-576,-566,-556,-547,-537,-528,-518,
+-509,-500,-490,-481,-472,-463,-454,-445,-436,-427,-418,-409,-400,-391,-383,-374,
+-365,-357,-348,-340,-331,-323,-314,-306,-298,-289,-281,-273,-265,-256,-248,-240,
+-232,-224,-216,-208,-200,-193,-185,-177,-169,-162,-154,-146,-139,-131,-123,-116,
+-108,-101,-93,-86,-79,-71,-64,-57,-49,-42,-35,-28,-21,-14,-7,0};
+
+class DSVoice : public OAKRA_Voice {
+public:
+	OAKRA_Module_OutputDS *driver;
+	OAKRA_Format format;
+	int formatShift;
+	IDirectSoundBuffer *ds_buf;
+	int buflen;
+	unsigned int cPlay;
+	int vol,pan;
+
+	virtual void setPan(int pan) {
+		this->pan = pan;
+		if(pan==0) ds_buf->SetPan(0);
+		else if(pan>0) ds_buf->SetPan(-voltbl[255-pan]);
+		else ds_buf->SetPan(voltbl[255+pan]);
+	}
+	virtual void setVol(int vol) {
+		this->vol = vol;
+		ds_buf->SetVolume(voltbl[vol]);
+	}
+	virtual int getVol() { return vol; }
+	virtual int getPan() { return pan; }
+	void setSource(OAKRA_Module *source) {
+		this->source = source;
+	}
+	virtual ~DSVoice() {
+		driver->freeVoiceInternal(this,true);
+		ds_buf->Release();
+	}
+	DSVoice(OAKRA_Module_OutputDS *driver, OAKRA_Format &format, IDirectSound *ds_dev, bool global) { 
+		this->driver = driver;
+		vol = 255;
+		pan = 0;
+		source = 0;
+		this->format = format;
+		formatShift = getFormatShift(format);
+		buflen = (format.rate * LATENCY_MS / 1000);
+
+		WAVEFORMATEX wfx;
+		memset(&wfx, 0, sizeof(wfx));
+		wfx.wFormatTag      = WAVE_FORMAT_PCM;
+		wfx.nChannels       = format.channels;
+		wfx.nSamplesPerSec  = format.rate;
+		wfx.nAvgBytesPerSec = format.rate * format.size;
+		wfx.nBlockAlign     = format.size;
+		wfx.wBitsPerSample  = (format.format==OAKRA_S16?16:8);
+		wfx.cbSize          = sizeof(wfx);
+
+		DSBUFFERDESC dsbd;
+		memset(&dsbd, 0, sizeof(dsbd));
+		dsbd.dwSize        = sizeof(dsbd);
+		dsbd.dwFlags       = DSBCAPS_GETCURRENTPOSITION2 | DSBCAPS_LOCSOFTWARE | DSBCAPS_CTRLVOLUME ;
+		if(global) dsbd.dwFlags |= DSBCAPS_GLOBALFOCUS ;
+		dsbd.dwBufferBytes = buflen  * format.size;
+		dsbd.lpwfxFormat   = &wfx;
+		
+		HRESULT hr = ds_dev->CreateSoundBuffer(&dsbd,&ds_buf,0);
+		cPlay = 0;
+
+		hr = ds_buf->Play(0,0,DSBPLAY_LOOPING);
+		int xxx=9;
+	}
+	
+	//not supported
+	virtual void volFade(int start, int end, int ms) {}
+
+	void update() {
+	    DWORD play, write;
+		HRESULT hr = ds_buf->GetCurrentPosition(&play, &write);
+		play >>= formatShift;
+		write >>= formatShift;
+		
+		int todo;
+		if(play<cPlay) todo = play + buflen - cPlay;
+		else todo = play - cPlay;
+
+		if(!todo) return;
+		
+
+		void* buffer1;
+		void* buffer2;
+		DWORD buffer1_length;
+		DWORD buffer2_length;
+		hr = ds_buf->Lock(
+			cPlay<<formatShift,todo<<formatShift,
+			&buffer1, &buffer1_length,
+			&buffer2, &buffer2_length,0
+			);
+
+		buffer1_length >>= formatShift;
+		buffer2_length >>= formatShift;
+		int done = 0;
+		if(source) {
+			done = source->generate(buffer1_length,buffer1);
+			if(done != buffer1_length) {
+				generateSilence(buffer1_length - done,(char *)buffer1 + (done<<formatShift),format.size);
+				generateSilence(buffer2_length,buffer2,format.size);
+				die();
+			} else {
+				if(buffer2_length) {
+					done = source->generate(buffer2_length,buffer2);
+					if(done != buffer2_length) {
+						generateSilence(buffer2_length - done,(char *)buffer2 + (done<<formatShift),format.size);
+						die();
+					}
+				}
+			}
+		}
+
+		ds_buf->Unlock(
+			buffer1, buffer1_length,
+			buffer2, buffer2_length);
+
+		cPlay = play;
+	}
+};
+
+class Data {
+public:
+	bool global;
+	IDirectSound* ds_dev;
+	std::vector<DSVoice *> voices;
+	CRITICAL_SECTION criticalSection;
+};
+
+
+class ThreadData {
+public:
+	ThreadData() { kill = dead = false; }
+	OAKRA_Module_OutputDS *ds;
+	bool kill,dead;
+};
+
+OAKRA_Module_OutputDS::OAKRA_Module_OutputDS() {
+	data = new Data();
+	((Data *)data)->global = false;
+	InitializeCriticalSection(&((Data *)data)->criticalSection);
+}
+
+OAKRA_Module_OutputDS::~OAKRA_Module_OutputDS() {
+	//ask the driver to shutdown, and wait for it to do so
+	((ThreadData *)threadData)->kill = true;
+	while(!((ThreadData *)threadData)->dead) Sleep(1);
+
+	////kill all the voices
+	std::vector<DSVoice *> voicesCopy = ((Data *)data)->voices;
+	int voices = (int)voicesCopy.size();
+	for(int i=0;i<voices;i++)
+		delete voicesCopy[i];
+
+	////free other resources
+	DeleteCriticalSection(&((Data *)data)->criticalSection);
+	((Data *)data)->ds_dev->Release();
+	delete (Data *)data;
+	delete (ThreadData *)threadData;
+}
+
+OAKRA_Voice *OAKRA_Module_OutputDS::getVoice(OAKRA_Format &format, OAKRA_Module *source) {
+	DSVoice *dsv = (DSVoice *)getVoice(format);
+	dsv->setSource(source);
+	return dsv;
+}
+
+OAKRA_Voice *OAKRA_Module_OutputDS::getVoice(OAKRA_Format &format) {
+	DSVoice *voice = new DSVoice(this,format,((Data *)data)->ds_dev,((Data *)data)->global);
+	((Data *)data)->voices.push_back(voice);
+	return voice;
+}
+void OAKRA_Module_OutputDS::freeVoice(OAKRA_Voice *voice) {
+	freeVoiceInternal(voice,false);
+}
+
+void OAKRA_Module_OutputDS::freeVoiceInternal(OAKRA_Voice *voice, bool internal) {
+	lock();
+	Data *data = (Data *)this->data;
+	int j = -1;
+	for(int i=0;i<(int)data->voices.size();i++)
+		if(data->voices[i] == voice) j = i;
+	if(j!=-1)
+		data->voices.erase(data->voices.begin()+j);
+	if(!internal)
+		delete voice;
+	unlock();
+}
+
+
+
+void OAKRA_Module_OutputDS::start(void *hwnd) {
+	HRESULT hr = CoInitialize(NULL);
+	IDirectSound* ds_dev;
+	hr = CoCreateInstance(CLSID_DirectSound,0,CLSCTX_INPROC_SERVER,IID_IDirectSound,(void**)&ds_dev);
+	
+	if(!hwnd) {
+		hwnd = GetDesktopWindow();
+		((Data *)data)->global = true;
+	}
+	
+	//use default device
+	hr = ds_dev->Initialize(0); 
+	hr = ds_dev->SetCooperativeLevel((HWND)hwnd, DSSCL_NORMAL);
+
+	((Data *)data)->ds_dev = ds_dev;
+}
+
+
+DWORD WINAPI updateProc(LPVOID lpParameter) {
+	ThreadData *data = (ThreadData *)lpParameter;
+	for(;;) {
+		if(data->kill) break;
+		data->ds->update();
+		Sleep(1);
+	}
+	data->dead = true;
+	return 0;
+}
+
+void OAKRA_Module_OutputDS::beginThread() {
+	DWORD updateThreadId;
+	threadData = new ThreadData();
+	((ThreadData *)threadData)->ds = this;
+	HANDLE updateThread = CreateThread(0,0,updateProc,threadData,0,&updateThreadId);
+	SetThreadPriority(updateThread,THREAD_PRIORITY_TIME_CRITICAL);
+	//SetThreadPriority(updateThread,THREAD_PRIORITY_HIGHEST);
+}
+
+void OAKRA_Module_OutputDS::endThread() {
+	((ThreadData *)threadData)->kill = true;
+}
+
+void OAKRA_Module_OutputDS::update() {
+	lock();
+	int voices = (int)((Data *)data)->voices.size();
+
+	//render all the voices
+	for(int i=0;i<voices;i++)
+		((Data *)data)->voices[i]->update();
+
+	//look for voices that are dead
+	std::vector<DSVoice *> deaders;
+	for(int i=0;i<voices;i++)
+		if(((Data *)data)->voices[i]->dead)
+			deaders.push_back(((Data *)data)->voices[i]);
+
+	//unlock the driver before killing voices!
+	//that way, the voice's death callback won't occur within the driver lock
+	unlock();
+
+	//kill those voices
+	for(int i=0;i<(int)deaders.size();i++) {
+		deaders[i]->callbackDied();
+		freeVoice(deaders[i]);
+	}
+}
+
+void OAKRA_Module_OutputDS::lock() {
+	EnterCriticalSection(  &((Data *)this->data)->criticalSection );
+}
+
+void OAKRA_Module_OutputDS::unlock() {
+	LeaveCriticalSection(  &((Data *)this->data)->criticalSection );
+}
+
diff --git a/src/drivers/win/input.cpp b/src/drivers/win/input.cpp
index 154e55e1..dcc3a407 100644
--- a/src/drivers/win/input.cpp
+++ b/src/drivers/win/input.cpp
@@ -125,6 +125,7 @@ static void UpdateTopRider(void);
 
 static uint32 JSreturn=0;
 int NoWaiting=0;
+bool turbo = false;
 
 #include "keyscan.h"
 static unsigned char *keys=0;
@@ -1226,8 +1227,8 @@ static struct
 	{ EMUCMD_FRAME_ADVANCE, 			SCAN_TAB, },
 	{ EMUCMD_SCREENSHOT,				SCAN_F9 },
 	{ EMUCMD_HIDE_MENU_TOGGLE,			SCAN_ESCAPE },
-	//{ EMUCMD_SPEED_SLOWER, 				SCAN_MINUS, }, // think about these
-	//{ EMUCMD_SPEED_FASTER, 				SCAN_EQUAL, }, // think about these
+	{ EMUCMD_SPEED_SLOWER, 				SCAN_MINUS, }, // think about these
+	{ EMUCMD_SPEED_FASTER, 				SCAN_EQUAL, }, // think about these
 	{ EMUCMD_SPEED_TURBO, 				SCAN_GRAVE, }, //tilde
 	{ EMUCMD_SAVE_SLOT_0, 				SCAN_0, },
 	{ EMUCMD_SAVE_SLOT_1, 				SCAN_1, },
@@ -1826,11 +1827,13 @@ void MapInput(void)
 
 void FCEUD_TurboOn(void)
 {
-	NoWaiting|=1;
+	//NoWaiting|=1;
+	turbo = 1;
 }
 
 void FCEUD_TurboOff(void)
 {
-	NoWaiting&=~1;
+	//NoWaiting&=~1;
+	turbo = false;
 }
 
diff --git a/src/drivers/win/main.cpp b/src/drivers/win/main.cpp
index d1c7603b..b4fa6465 100644
--- a/src/drivers/win/main.cpp
+++ b/src/drivers/win/main.cpp
@@ -490,18 +490,17 @@ doloopy:
 	{
 		while(GameInfo)
 		{
-	        uint8 *gfx=0;
-			int32 *sound=0;
-			int32 ssize=0;
+	        uint8 *gfx=0; ///contains framebuffer
+			int32 *sound=0; ///contains sound data buffer
+			int32 ssize=0; ///contains sound samples count
 
 			#ifdef _USE_SHARED_MEMORY_
 			UpdateBasicBot();
 			#endif
 			FCEU_UpdateBot();
 
-			FCEUI_Emulate(&gfx, &sound, &ssize, 0);
-			xbsave = gfx;
-			FCEUD_Update(gfx, sound, ssize);
+			FCEUI_Emulate(&gfx, &sound, &ssize, 0); //emulate a single frame
+			FCEUD_Update(gfx, sound, ssize); //update displays and debug tools
 
 			 //mbg 6/30/06 - close game if we were commanded to by calls nested in FCEUI_Emulate()
 			 if(closeGame)
@@ -510,7 +509,21 @@ doloopy:
 				GameInfo = 0;
 			 }
 
-			
+		}
+		//xbsave = NULL;
+		RedrawWindow(hAppWnd,0,0,RDW_ERASE|RDW_INVALIDATE);
+		StopSound();
+	}
+	Sleep(50);
+	if(!exiting)
+		goto doloopy;
+
+	doexito:
+	DriverKill();
+	timeEndPeriod(1);
+	FCEUI_Kill();
+	return(0);
+}
 
 			//mbg merge 7/19/06 
 			//--------this code was added by tasbuild
@@ -549,21 +562,8 @@ doloopy:
 				stopCount=0;
 			}*/
 			//-----------------------------------
-		}
-		xbsave = NULL;
-		RedrawWindow(hAppWnd,0,0,RDW_ERASE|RDW_INVALIDATE);
-		StopSound();
-	}
-	Sleep(50);
-	if(!exiting)
-		goto doloopy;
 
-	doexito:
-	DriverKill();
-	timeEndPeriod(1);
-	FCEUI_Kill();
-	return(0);
-}
+
 
 //mbg merge 7/19/06 - the function that contains the code that used to just be UpdateFCEUWindow() and FCEUD_UpdateInput()
 void _updateWindow() {
@@ -780,15 +780,10 @@ void _updateWindow() {
 
 void FCEUD_Update(uint8 *XBuf, int32 *Buffer, int Count)
 {
-	static int skipcount = 0;
-	int temp_fps_scale=(NoWaiting&1)?(256*16):fps_scale;
-	int maxskip = (temp_fps_scale<=256) ? 0 : temp_fps_scale>>8;
-
-	int ocount = Count;
+	//mbg merge 7/19/06 - leaving this untouched but untested
+	/*int ocount = Count;
 	// apply frame scaling to Count
 	Count = (Count<<8)/temp_fps_scale;
-
-	//mbg merge 7/19/06 - leaving this untouched but untested
 	//Disable sound and throttling for BotMode--we want max speed!
 	if(FCEU_BotMode())
 	{
@@ -804,19 +799,41 @@ void FCEUD_Update(uint8 *XBuf, int32 *Buffer, int Count)
 		UpdateFCEUWindow();
 		FCEUD_UpdateInput();
 		return;
-	}
+	}*/
 
-	//mbg naive code
+	extern bool turbo; //hack
+
+	//write all the sound we generated.
 	if(soundo && Buffer && Count) {
-		int32 writeSize = GetWriteSound();
-		int32 writeCount = Count;
-		FCEUD_WriteSoundData(Buffer,temp_fps_scale,MAX(writeSize,writeCount));
+		void FCEUD_WriteSoundData_new(int32 *Buffer, int scale, int Count, bool turbo);
+		FCEUD_WriteSoundData_new(Buffer,fps_scale,Count,turbo);
+		//FCEUD_WriteSoundData(Buffer,fps_scale,Count);
 	}
 
+	//blit the framebuffer
 	if(XBuf)
 		FCEUD_BlitScreen(XBuf);
+
+	//update debugging displays
 	_updateWindow();
 
+
+	//throttle
+	bool skip = false;
+	for(;;) {
+		if(!(eoptions&EO_NOTHROTTLE)) //if throttling is enabled..
+			if(!turbo) //and turbo is disabled..
+				if(!FCEUI_EmulationPaused()) //and we're not paused..
+					//if(SpeedThrottle()) {//...then check whether we need to throttle
+					//	//idle..
+					//	Sleep(0);
+					//	continue;
+					//}
+				{}
+		break;
+	}
+
+
 	//delay until we unpause. we will only stick here if we're paused by a breakpoint or debug command
 	while(FCEUI_EmulationPaused() && inDebugger)
 	{
diff --git a/src/drivers/win/oakra.h b/src/drivers/win/oakra.h
new file mode 100644
index 00000000..5e815bf4
--- /dev/null
+++ b/src/drivers/win/oakra.h
@@ -0,0 +1,209 @@
+//this code is sloppily ripped from an unfinished sound system written for internal use by m.gambrell
+//it is released into the public domain and its stability is not warranted
+
+
+#define _USE_MATH_DEFINES
+#include <math.h>
+#include <stdlib.h>
+#include <map>
+#include <stack>
+#include <ctype.h>
+#include <queue>
+
+
+const int OAKRA_U8 = 0;
+const int OAKRA_S16 = 1;
+const int OAKRA_S32 = 2;
+
+const int OAKRA_STATUS_STOPPED = 0;
+const int OAKRA_STATUS_PLAYING = 1;
+
+struct OAKRA_Format {
+	int channels, format, rate, size;
+};
+
+//implements an easy to use, bland callback
+template <typename T> struct OAKRA_Callback {
+	OAKRA_Callback() { func = 0; param = 0; }
+	T operator()() { if(func) return func(param); }
+	T (*func)(void *param);
+	void *param;
+};
+template <typename T, typename TARG> struct OAKRA_ArgumentCallback {
+	OAKRA_ArgumentCallback() { func = 0; param = 0; }
+	T operator()(TARG arg) { if(func) return func(param,arg); }
+	T (*func)(void *param, TARG arg);
+	void *param;
+};
+
+class OAKRA_Module {
+public:
+	virtual int generate(int samples, void *buf) { return 0; }
+
+	int adapt_to_2S16(int samples, void *buf, OAKRA_Format &sourceFormat) {
+		short *sbuf = (short *)buf;
+		unsigned char *bbuf = (unsigned char *)buf;
+		if(sourceFormat.format == OAKRA_S16) {
+			if(sourceFormat.channels == 2) return samples;
+			for(int i=samples-1,j=samples*2-2;i>=0;i--,j-=2)
+				sbuf[j] = sbuf[j+1] = sbuf[i];
+		} else {
+			if(sourceFormat.channels == 1)
+				for(int i=samples-1,j=samples*2-2;i>=0;i--,j-=2)
+					sbuf[j] = sbuf[j+1] = ((int)bbuf[i]-128)<<8;
+			else
+				for(int j=samples*2-2;j>=0;j--)
+					sbuf[j] = sbuf[j+1] = ((int)bbuf[j]-128)<<8;
+		}
+		return samples;
+	}
+
+	//trashes some samples using the buffer provided
+	void trash(int samples, void *buf, int bufsamples) {
+		while(samples) {
+			int todo = std::min(samples,bufsamples);
+			generate(todo,buf);
+			samples -= todo;
+		}
+	}
+
+	static int calcSize(OAKRA_Format &format) {
+		int size = format.channels;
+		if(format.format == OAKRA_S16) size *= 2;
+		return size;
+	}
+
+	static int getFormatShift(OAKRA_Format &format) {
+		if(format.size==1) return 0;
+		else if(format.size==2) return 1;
+		else if(format.size==4) return 2;
+		return -1; //try and crash!
+	}
+
+	static void *malloc(int len) { 
+		void *ptr = ::malloc(len);
+		return ptr;
+	}
+	void *malloc(int sampsize, int len) { return malloc(sampsize*len); }
+	void *malloc(int sampsize, int channels, int len) { return malloc(sampsize*len*channels); }
+	void *realloc(void *ptr, int len) {
+		ptr = ::realloc(ptr,len);
+		return ptr;
+	}
+	void free(void *ptr) {
+		::free(ptr);
+	}
+	
+
+	static int generateNoise(int samples, void *buf, int sampleSize) {
+		for(int i=0;i<samples;i++)
+			for(int j=0;j<sampleSize;j++)
+				((char *)buf)[i*sampleSize+j] = rand();
+		return samples;
+	}
+
+	static int generateSilence(int samples, void *buf, int sampleSize) {
+		char *cbuf = (char *)buf;
+		int n = sampleSize*samples;
+		memset(buf,0,n);
+		return samples;
+	}
+
+	static int generateSilence(int samples, void *buf, OAKRA_Format &format) {
+		char *cbuf = (char *)buf;
+		int n = format.size*samples;
+		memset(buf,0,n);
+		return samples;
+	}
+
+
+	template<typename T>
+	T *malloc() {
+		return (T *)malloc(sizeof(T));
+	}
+	template<typename T>
+	T *malloc(int amt) {
+		return (T *)malloc(sizeof(T)*amt);
+	}
+	template<typename T>
+	T **nmalloc(int number) {
+		return (T **)malloc(sizeof(T)*number);
+	}
+};
+
+class OAKRA_IQueryFormat {
+public:
+	virtual OAKRA_Format &queryFormat() = 0;
+};
+
+//this is basically a filter class
+//one source, one sink
+class OAKRA_BasicModule : public OAKRA_Module {
+public:
+	OAKRA_Module *source, *sink;
+	OAKRA_BasicModule() { source = sink = 0; }
+};
+
+class OAKRA_Voice : public OAKRA_BasicModule {
+public:
+	OAKRA_Voice() { dead = false; }
+	virtual ~OAKRA_Voice() { }
+	virtual void setPan(int pan) = 0;
+	virtual int getPan() =0;
+	virtual void setVol(int vol) = 0;
+	virtual int getVol()=0;
+	virtual void setSource(OAKRA_Module *source) = 0;
+	
+	virtual void volFade(int start, int end, int ms)=0;
+
+	//call this when youre in the middle of rendering the voice, but have decided to quit
+	//the driver will then trash the voice as soon as it gets the chance.
+	//you dont have to have the driver locked to call it. that would be illogical,
+	//as the driver is currently locked while it is rendering!
+	virtual void die() {
+		dead = true;
+	}
+
+	//indicates whether a voice is dead
+	bool dead;
+
+	//callback fired when a voice dies
+	OAKRA_Callback<void> callbackDied;
+};
+
+
+class OAKRA_OutputDriver : public OAKRA_BasicModule {
+public:
+	virtual ~OAKRA_OutputDriver() {} ;
+    virtual OAKRA_Voice *getVoice(OAKRA_Format &format) = 0;
+	virtual OAKRA_Voice *getVoice(OAKRA_Format &format, OAKRA_Module *source) = 0;
+
+	//this should be safe to call from within a driver callback. in general, nothing else will be.
+	//even if you dont delete the voice (if youre recycling it) you should clear out the source asap
+	virtual void freeVoice(OAKRA_Voice *voice) = 0;
+
+	virtual void lock() = 0;
+	virtual void unlock() = 0;
+};
+
+
+class OAKRA_Module_OutputDS : public OAKRA_OutputDriver {
+	
+	void *threadData;
+	void *data;
+
+public:
+
+	OAKRA_Module_OutputDS();
+	~OAKRA_Module_OutputDS();
+	OAKRA_Voice *getVoice(OAKRA_Format &format);
+	OAKRA_Voice *getVoice(OAKRA_Format &format, OAKRA_Module *source);
+	void freeVoice(OAKRA_Voice *voice);
+	void freeVoiceInternal(OAKRA_Voice *voice, bool internal);
+	void start(void *hwnd);
+	void update();
+	void beginThread();
+	void endThread();
+	void lock();
+	void unlock();
+};
diff --git a/src/drivers/win/sound.cpp b/src/drivers/win/sound.cpp
index 50ce44a0..85c4ea85 100644
--- a/src/drivers/win/sound.cpp
+++ b/src/drivers/win/sound.cpp
@@ -18,6 +18,7 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
+#include <list>
 
 LPDIRECTSOUND ppDS=0;           /* DirectSound object. */
 LPDIRECTSOUNDBUFFER ppbuf=0;    /* Primary buffer object. */
@@ -131,6 +132,10 @@ int32 GetMaxSound(void)
  return( BufHowMuch >> bittage);
 }
 
+///enqueues the given samples for playback
+static void EnqueueSamples(void *data, uint32 len) {
+}
+
 static int RawWrite(void *data, uint32 len)
 {
  //uint32 cw; //mbg merge 7/17/06 removed
@@ -156,11 +161,12 @@ static int RawWrite(void *data, uint32 len)
 
 
   // THIS LIMITS THE EMULATION SPEED
-  if((!NoWaiting) || (soundoptions&SO_OLDUP))
-  while(!(curlen=RawCanWrite()))
-  {   
-   Sleep(1);
-  }
+  //if((!NoWaiting) || (soundoptions&SO_OLDUP))
+  //printf("RawCanWrite(): %d\n",RawCanWrite());
+  //while(!(curlen=RawCanWrite()))
+  //{  
+  // Sleep(1);
+  //}
 
   if(curlen>len) curlen=len;
 
@@ -211,8 +217,8 @@ static int RawWrite(void *data, uint32 len)
   len-=curlen;
   data=(uint8 *)data+curlen; //mbg merge 7/17/06 reworked to be type proper
 
-  if(len && !NoWaiting && (fps_scale <= 256 || (soundoptions&SO_OLDUP)))
-   Sleep(1); // do some extra sleeping if we think there's time and we're not scaling up the FPS or in turbo mode
+  //if(len && !NoWaiting && (fps_scale <= 256 || (soundoptions&SO_OLDUP)))
+  // Sleep(1); // do some extra sleeping if we think there's time and we're not scaling up the FPS or in turbo mode
 
  } // end while(len) loop
 
@@ -222,6 +228,212 @@ static int RawWrite(void *data, uint32 len)
 
 int silencer=0;
 
+#undef min
+#undef max
+#include "oakra.h"
+
+OAKRA_Module_OutputDS *dsout;
+
+class BufferSet {
+public:
+
+	static const int BufferSize = 1024;
+	static const int BufferSizeBits = 10;
+	static const int BufferSizeBitmask = 1023;
+	
+	class Buffer {
+	public:
+		int decay, size, length;
+		Buffer(int size) { length = 0; this->size = size; data = OAKRA_Module::malloc(size); }
+		int getRemaining() { return size-length; }
+		void *data;
+		~Buffer() { delete data; }
+	};
+
+	std::vector<Buffer*> liveBuffers;
+	std::vector<Buffer*> freeBuffers;
+	int length;
+	int offset; //offset of beginning of addressing into current buffer
+	int bufferCount;
+
+	BufferSet() {
+		offset = length = bufferCount = 0;
+	}
+
+	//causes the oldest free buffer to decay one unit. kills it if it gets too old
+	void decay(int threshold) {
+		if(freeBuffers.empty()) return;
+		if(freeBuffers[0]->decay++>=threshold) {
+			delete freeBuffers[0];
+			freeBuffers.erase(freeBuffers.begin());
+		}
+	}
+
+	Buffer *getBuffer() {
+		//try to get a buffer from the free pool first
+		//if theres nothing in the free pool, get a new buffer
+		if(!freeBuffers.size()) return getNewBuffer();
+		//otherwise, return the last thing in the free pool (most recently freed)
+		Buffer *ret = *--freeBuffers.end();
+		freeBuffers.erase(--freeBuffers.end());
+		return ret;
+	}
+
+	//adds the buffer to the free list
+	void freeBuffer(Buffer *buf) {
+		freeBuffers.push_back(buf);
+		buf->decay = 0;
+		buf->length = 0;
+	}
+
+	Buffer *getNewBuffer() {
+		bufferCount++;
+		return new Buffer(BufferSize);
+	}
+
+	short getShortAtIndex(int addr) {
+		addr <<= 1; //shorts are 2bytes
+		int buffer = (addr+offset)>>BufferSizeBits;
+		int ofs = (addr+offset) & BufferSizeBitmask;
+		return *(short*)((char*)liveBuffers[buffer]->data+ofs);
+	}
+
+	//dequeues the specified number of bytes
+	void dequeue(int length) {
+		offset += length;
+		while(offset >= BufferSize) {
+			Buffer *front = liveBuffers[0];
+			freeBuffer(front);
+			liveBuffers.erase(liveBuffers.begin());
+			offset -= BufferSize;
+		}
+		this->length -= length;
+	}
+
+
+	void enqueue(int length, void *data) {
+		int todo = length;
+		int done = 0;
+
+		//if there are no buffers in the queue, then we definitely need one before we start
+		if(liveBuffers.empty()) liveBuffers.push_back(getBuffer());
+		
+		while(todo) {
+			//check the frontmost buffer
+			Buffer *end = *--liveBuffers.end();
+			int available = std::min(todo,end->getRemaining());
+			memcpy((char*)end->data + end->length,(char*)data + done,available);
+			end->length += available;
+			todo -= available;
+			done += available;
+			
+			//we're going to need another buffer
+			if(todo != 0)
+				liveBuffers.push_back(getBuffer());
+		}
+
+		this->length += length;
+	}
+};
+
+
+class Player : public OAKRA_Module {
+public:
+
+	BufferSet buffers;
+	int cursor;
+	
+	bool turbo;
+	int scale;
+
+	//not interpolating! someday it will! 
+	int generate(int samples, void *buf) {
+
+		int incr = 256;
+		int bufferSamples = buffers.length>>1;
+
+		//if we're we're too far behind, playback faster
+		if(bufferSamples > 44100*3/60) {
+			int behind = bufferSamples - 44100/60;
+			incr = behind*256*60/44100/2;
+			//we multiply our playback rate by 1/2 the number of frames we're behind
+		}
+		if(incr<256) printf("OHNO -- %d -- shouldnt be less than 256!\n",incr); //sanity check: should never be less than 256
+
+		incr = (incr*scale)>>8; //apply scaling factor
+
+		//figure out how many dest samples we can generate at this rate without running out of source samples
+		int destSamplesCanGenerate = (bufferSamples<<8) / incr;
+
+		int todo = std::min(destSamplesCanGenerate,samples);
+		short *sbuf = (short*)buf;
+		for(int i=0;i<todo;i++) {
+			sbuf[i] = buffers.getShortAtIndex(cursor>>8);
+			cursor += incr;
+		}
+		buffers.dequeue((cursor>>8)<<1);
+		cursor &= 255;
+		memset(sbuf+todo,0,(samples-todo)<<1);
+		return samples;
+	}
+
+	Player() {
+		scale = 256;
+		cursor = 0;
+		turbo = false;
+	}
+
+	///receives and enqueues s16 stereo samples
+	void receive(int samples, short *buf, bool turbo, int scale) {
+
+		dsout->lock();
+		buffers.enqueue(samples*2,buf);
+		buffers.decay(60);
+		this->scale = scale;
+		dsout->unlock();
+
+		//throttle
+//		//wait for the buffer to be satisfactorily low before continuing
+		if(!turbo) {
+			for(;;) {
+				dsout->lock();
+				int remain = buffers.length>>1;
+				dsout->unlock();
+				if(remain<44100/60) break;
+				Sleep(1);
+			}
+		}
+	}
+
+};
+Player *player;
+
+
+void sound_init_new() {
+	dsout = new OAKRA_Module_OutputDS();
+	dsout->start(0);
+	dsout->beginThread();
+	OAKRA_Format fmt;
+	fmt.format = OAKRA_S16;
+	fmt.channels = 1;
+	fmt.rate = 44100;
+	fmt.size = 2;
+	OAKRA_Voice *voice = dsout->getVoice(fmt);
+	player = new Player();
+	dsout->lock();
+	voice->setSource(player);
+	dsout->unlock();
+}
+
+void FCEUD_WriteSoundData_new(int32 *Buffer, int scale, int Count, bool turbo) {
+	#define WSD_BUFSIZE (2 * 96000 / 50)
+	static int16 MBuffer[WSD_BUFSIZE*2];
+
+	for(int i=0;i<Count;i++)
+		MBuffer[i] = Buffer[i];
+	player->receive(Count,MBuffer,turbo,scale);
+}
+
 int FCEUD_WriteSoundData(int32 *Buffer, int scale, int Count)
 {
 #define WSD_BUFSIZE (2 * 96000 / 50)
@@ -230,21 +442,25 @@ int FCEUD_WriteSoundData(int32 *Buffer, int scale, int Count)
 	int iCount=0;
 	static int16 MBuffer[WSD_BUFSIZE*2];
 
-	if(!(soundoptions&SO_OLDUP))
-	{
-		if(FCEUI_EmulationPaused())
-			memset(MBuffer, 0, WSD_BUFSIZE); // slow and/or unnecessary
+	//if(!(soundoptions&SO_OLDUP))
+	//{
+	//	if(FCEUI_EmulationPaused())
+	//		memset(MBuffer, 0, WSD_BUFSIZE); // slow and/or unnecessary
 
-		if(FCEUI_EmulationPaused()) scale >>= 1;
+	//	if(FCEUI_EmulationPaused()) scale >>= 1;
 
-		// limit frequency change to between 50% and 200%
-		if(scale > 512) scale = 512;
-		if(scale < 128) scale = 128;
-	}
+	//	// limit frequency change to between 50% and 200%
+	//	if(scale > 512) scale = 512;
+	//	if(scale < 128) scale = 128;
+	//}
 
 //	for(;Count>0;Count-=WSD_BUFSIZE)
 	{
-		int amt = (soundoptions&SO_OLDUP) ? Count : (Count > WSD_BUFSIZE ? WSD_BUFSIZE : Count);
+		//int amt = (soundoptions&SO_OLDUP) ? Count : (Count > WSD_BUFSIZE ? WSD_BUFSIZE : Count);
+		int amt = Count;
+
+		int cando = RawCanWrite();
+		printf("Count/cando %d/%d\n",amt,cando);
 
 		if(!bittage)
 		{
@@ -258,7 +474,7 @@ int FCEUD_WriteSoundData(int32 *Buffer, int scale, int Count)
 				for(P=0;P<amt;P++)
 					*(((uint8*)MBuffer)+P)=((int8)(Buffer[P*scale/256]>>8))^128;
 
-			RawWrite(MBuffer,amt);
+			RawWrite(MBuffer,cando);
 		}
 		else // force 8-bit sound is off:
 		{
@@ -272,7 +488,7 @@ int FCEUD_WriteSoundData(int32 *Buffer, int scale, int Count)
 				for(P=0;P<amt;P++)
 					MBuffer[P]=Buffer[P*scale/256];
 
-			RawWrite(MBuffer,amt * 2);
+			RawWrite(MBuffer,cando * 2);
 		}
 
 		iCount+=amt;
@@ -364,6 +580,8 @@ int InitSound()
  DSCAPS dscaps;
  DSBCAPS dsbcaps;
 
+ sound_init_new();
+
  memset(&wf,0x00,sizeof(wf));
  wf.wFormatTag = WAVE_FORMAT_PCM;
  wf.nChannels = 1;
@@ -744,4 +962,7 @@ void FCEUD_SoundVolumeAdjust(int n)
 	FCEU_DispMessage("Sound volume %d.", soundvolume);
 }
 
+//-----------
+
+
 #include "wave.cpp"
diff --git a/src/drivers/win/throttle.cpp b/src/drivers/win/throttle.cpp
index 443194b4..2303bbc4 100644
--- a/src/drivers/win/throttle.cpp
+++ b/src/drivers/win/throttle.cpp
@@ -18,6 +18,8 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
+//http://www.geisswerks.com/ryan/FAQS/timing.html
+
 static uint64 tmethod,tfreq;
 static uint64 desiredfps;
 
@@ -27,8 +29,11 @@ int32 fps_scale = 256;
 
 static void RefreshThrottleFPS(void)
 {
+	printf("WTF\n");
+	fflush(stdout);
  desiredfps=FCEUI_GetDesiredFPS()>>8;
  desiredfps=(desiredfps*fps_scale)>>8;
+ 
 }
 
 static uint64 GetCurTime(void)
@@ -47,56 +52,102 @@ static uint64 GetCurTime(void)
 
 }
 
+static uint64 ttime,ltime;
+
 static void InitSpeedThrottle(void)
 {
- tmethod=0;
- if(QueryPerformanceFrequency((LARGE_INTEGER*)&tfreq))
- {
-  tmethod=1;
- }
- else
-  tfreq=1000;
- tfreq<<=16;    /* Adjustment for fps returned from FCEUI_GetDesiredFPS(). */
+	timeBeginPeriod(1);
+	SetThreadAffinityMask(GetCurrentThread(),1);
+	
+
+	tmethod=0;
+	if(QueryPerformanceFrequency((LARGE_INTEGER*)&tfreq)) {
+		tmethod=1;
+	}
+	else tfreq=1000;
+
+	tfreq<<=16;    /* Adjustment for fps returned from FCEUI_GetDesiredFPS(). */
+	ltime = 0; //mbg
+}
+
+///Resets the throttle timing. use this when the player releases the turbo
+void ResetSpeedThrottle() {
+	ltime = 0;
 }
 
 
 static int SpeedThrottle(void)
 {
- static uint64 ttime,ltime;
+	//the desired running time for this frame
+	uint64 desiredRunningTime = tfreq/desiredfps;
 
- waiter:
+	ttime = GetCurTime();
 
- ttime=GetCurTime();
+	//if this is our first time, save ltime and bail out
+	if(ltime == 0) {
+		ltime = ttime;
+		return 0;
+	}
 
- if( (ttime-ltime) < (tfreq/desiredfps) )
- {
-  uint64 sleepy;
-  sleepy=(tfreq/desiredfps)-(ttime-ltime);  
-  sleepy*=1000;
-  if(tfreq>=65536)
-	  sleepy/=tfreq>>16;
-  else
-      sleepy=0;
-  if(sleepy>100)
-  {
-   // block for a max of 100ms to
-   // keep the gui responsive
-   Sleep(100);
-   return 1;
-  }
-  Sleep(sleepy);
-  goto waiter;
- }
- if( (ttime-ltime) >= (tfreq*4/desiredfps))
-  ltime=ttime;
- else
- {
-  ltime+=tfreq/desiredfps;
+	//otherwise calculate our delta
+	uint64 delta = ttime-ltime;
+	
+	/*printf("%20I64d %20I64d\n",delta,desiredRunningTime);
+	fflush(stdout);*/
+	if( delta < desiredRunningTime ) {
+		//if the elapsed time is less than the desired running time
 
-  if( (ttime-ltime) >= (tfreq/desiredfps) ) // Oops, we're behind!
-  return(1);
- }
- return(0);
+		//sleepy gets the time that needs to be slept.
+		//it is then converted to ms<<16
+		uint64 sleepy = desiredRunningTime-delta; 
+		//sleepy *= 1000;
+		//if we have more than 1 ms to sleep, sleep 1 ms and return
+		if(sleepy>=65536) {
+			Sleep(1);
+			return 1;
+		} else {
+			//otherwise, we can't throttle for less than 1ms. assume we're close enough and return
+			ltime += desiredRunningTime;
+			return 0;
+		}
+
+		//sleepy*=1000;
+		//if(tfreq>=65536)
+		//	sleepy/=tfreq>>16;
+		//else
+		//	sleepy=0;
+		//if(sleepy>100)
+		//{
+		//	// block for a max of 100ms to
+		//	// keep the gui responsive
+		//	Sleep(100);
+		//	return 1;
+		//}
+		//Sleep(sleepy);
+		//goto waiter;
+	} else {
+		//we're behind...
+		if(delta > 2*desiredRunningTime) {
+			//if we're behind by 2 frames, then reset the throttling
+			ResetSpeedThrottle();
+		} else {
+			//we're only behind by part of a frame
+			//just don't throttle!
+			ltime += desiredRunningTime;
+		}
+		return 0;
+	}
+	//}
+	//if( (ttime-ltime) >= (tfreq*4/desiredfps))
+	//	ltime=ttime;
+	//else
+	//{
+	//	ltime+=tfreq/desiredfps;
+
+	//	if( (ttime-ltime) >= (tfreq/desiredfps) ) // Oops, we're behind!
+	//		return(1);
+	//}
+	//return(0);
 }
 
 // Quick code for internal FPS display.
diff --git a/src/fceu.cpp b/src/fceu.cpp
index 682949a0..bdee8934 100644
--- a/src/fceu.cpp
+++ b/src/fceu.cpp
@@ -426,6 +426,9 @@ void AutoFire(void)
 
 void UpdateRewind(void);
 
+///Emulates a single frame.
+
+///Skip may be passed in, if FRAMESKIP is #defined, to cause this to emulate more than one frame
 void FCEUI_Emulate(uint8 **pXBuf, int32 **SoundBuf, int32 *SoundBufSize, int skip)
 {
  int r,ssize;
@@ -452,7 +455,7 @@ void FCEUI_Emulate(uint8 **pXBuf, int32 **SoundBuf, int32 *SoundBufSize, int ski
  FCEU_UpdateInput();
  if(geniestage!=1) FCEU_ApplyPeriodicCheats();
  r=FCEUPPU_Loop(skip);
-
+ 
  ssize=FlushEmulateSound();
 
 //#ifdef WIN32
diff --git a/src/types.h b/src/types.h
index 7fc01e4a..f2b14624 100644
--- a/src/types.h
+++ b/src/types.h
@@ -53,6 +53,7 @@ typedef signed int int32;
 #define fstat _fstat
 #define mkdir _mkdir
 #define alloca _alloca
+#define snprintf _snprintf
 #define W_OK 2
 #define R_OK 2
 #define X_OK 1
diff --git a/vc7/fceu.vcproj b/vc7/fceu.vcproj
index 063aabf6..33c22a20 100644
--- a/vc7/fceu.vcproj
+++ b/vc7/fceu.vcproj
@@ -594,6 +594,9 @@
 				<File
 					RelativePath="..\src\drivers\win\ntview.h">
 				</File>
+				<File
+					RelativePath="..\src\drivers\win\OutputDS.cpp">
+				</File>
 				<File
 					RelativePath="..\src\drivers\win\ppuview.cpp">
 				</File>
diff --git a/vc8/fceux.vcproj b/vc8/fceux.vcproj
index 7aa53111..a1dd3fc8 100644
--- a/vc8/fceux.vcproj
+++ b/vc8/fceux.vcproj
@@ -70,7 +70,7 @@
 				AdditionalDependencies="comctl32.lib vfw32.lib dxguid.lib winmm.lib dinput.lib ws2_32.lib ddraw.lib dsound.lib"
 				LinkIncremental="2"
 				GenerateDebugInformation="true"
-				SubSystem="2"
+				SubSystem="1"
 				EntryPointSymbol="mainCRTStartup"
 				TargetMachine="1"
 			/>
@@ -211,12 +211,12 @@
 				EnableIntrinsicFunctions="true"
 				FavorSizeOrSpeed="1"
 				OmitFramePointers="true"
-				AdditionalIncludeDirectories="../zlib"
+				AdditionalIncludeDirectories="../src/drivers/win/zlib"
 				PreprocessorDefinitions="WIN32;NDEBUG;MSVC;_CRT_SECURE_NO_DEPRECATE;_WIN32_WINDOWS=0x0410;WINVER=0x0410;NETWORK;LSB_FIRST;_USE_32BIT_TIME_T;FCEUDEF_DEBUGGER;_USE_SHARED_MEMORY_"
 				RuntimeLibrary="0"
 				UsePrecompiledHeader="0"
 				WarningLevel="3"
-				Detect64BitPortabilityProblems="true"
+				Detect64BitPortabilityProblems="false"
 				DebugInformationFormat="3"
 			/>
 			<Tool
@@ -634,6 +634,14 @@
 					RelativePath="..\src\drivers\common\config.h"
 					>
 				</File>
+				<File
+					RelativePath="..\src\drivers\common\configSys.cpp"
+					>
+				</File>
+				<File
+					RelativePath="..\src\drivers\common\configSys.h"
+					>
+				</File>
 				<File
 					RelativePath="..\src\drivers\common\hq2x.cpp"
 					>
@@ -686,10 +694,6 @@
 			<Filter
 				Name="win"
 				>
-				<File
-					RelativePath="..\src\asm.h"
-					>
-				</File>
 				<File
 					RelativePath="..\src\drivers\win\aviout.cpp"
 					>
@@ -698,6 +702,10 @@
 					RelativePath="..\src\drivers\win\basicbot.cpp"
 					>
 				</File>
+				<File
+					RelativePath="..\src\drivers\win\basicbot.h"
+					>
+				</File>
 				<File
 					RelativePath="..\src\drivers\win\cdlogger.cpp"
 					>
@@ -914,6 +922,14 @@
 					RelativePath="..\src\drivers\win\ntview.h"
 					>
 				</File>
+				<File
+					RelativePath="..\src\drivers\win\oakra.h"
+					>
+				</File>
+				<File
+					RelativePath="..\src\drivers\win\OutputDS.cpp"
+					>
+				</File>
 				<File
 					RelativePath="..\src\drivers\win\ppuview.cpp"
 					>