diff --git a/README.md b/README.md
index 5720b2f..046e573 100644
--- a/README.md
+++ b/README.md
@@ -45,17 +45,27 @@ $ ./build/darwin-amd64/mettaur XXXX.gba
 
 ## ToDo
 
-- [ ] Sound DMA
-- [ ] Clear sound
 - [ ] Window
 - [ ] Mosaic
 - [ ] Blend
 - [ ] GUI
 - [ ] Serial communication
 - [ ] BG mode5
+- [ ] GameBoy Compatibility
 - [ ] Debug feature
 - [ ] Fix some bugs
 
+## Game Compatibility List
+
+| Game Title             | Compatibility      |
+| -------------------- | ------------- |
+| バトルネットワーク ロックマンエグゼ3 BLACK | ✅ |
+| ロックマンエグゼ4 トーナメント ブルームーン | ✅ |
+| ロックマンエグゼ6 電脳獣グレイガ・電脳獣ファルザー | ✅ |
+| ポケットモンスター ファイアレッド | ✅ |
+| ドラゴンクエストモンスターズ キャラバンハート | 🔺 |
+
+
 ## References
 
 - [GBATEK](https://problemkaputt.de/gbatek.htm)
diff --git a/pkg/gba/apu.go b/pkg/gba/apu.go
index 8865657..ae37272 100644
--- a/pkg/gba/apu.go
+++ b/pkg/gba/apu.go
@@ -18,7 +18,7 @@ const (
 	BUFF_SAMPLES_MSK         = ((BUFF_SAMPLES) - 1)
 	SAMPLE_TIME      float64 = 1.0 / SND_FREQUENCY
 	SAMPLE_RATE              = SND_FREQUENCY
-	STREAM_LEN               = 2940 // 2 * 2 * sampleRate * (1/60)
+	STREAM_LEN               = 2184 // 2 * 2 * sampleRate * (1/60)
 )
 
 const (
@@ -498,7 +498,7 @@ func (g *GBA) soundClock(cycles uint32) {
 
 	cnth := uint16(g._getRAM(ram.SOUNDCNT_H)) // snd_pcm_vol
 	volADiv, volBDiv := int16((cnth>>2)&0b1), int16((cnth>>3)&0b1)
-	sampCh4, sampCh5 := fifoASamp>>volADiv, fifoBSamp>>volBDiv
+	sampCh4, sampCh5 := (int16(fifoASamp)<<1)>>volADiv, (int16(fifoBSamp)<<1)>>volBDiv
 	// sampCh4, sampCh5 = 0, 0
 
 	// Left
diff --git a/pkg/gba/dma.go b/pkg/gba/dma.go
index 58392b3..462dcd5 100644
--- a/pkg/gba/dma.go
+++ b/pkg/gba/dma.go
@@ -14,11 +14,14 @@ const (
 )
 
 type DMA struct {
-	io       [12]byte
-	src, dst uint32
+	io                  [12]byte
+	src, dst            uint32
+	count, defaultCount int
 }
 
-func NewDMA() [4]*DMA       { return [4]*DMA{&DMA{}, &DMA{}, &DMA{}, &DMA{}} }
+func NewDMA() [4]*DMA {
+	return [4]*DMA{&DMA{defaultCount: 0x4000}, &DMA{defaultCount: 0x4000}, &DMA{defaultCount: 0x4000}, &DMA{defaultCount: 0x10000}}
+}
 func (ch *DMA) cnt() uint32 { return util.LE32(ch.io[8:]) }
 func (ch *DMA) setCnt(v uint32) {
 	ch.io[8], ch.io[9], ch.io[10], ch.io[11] = byte(v), byte(v>>8), byte(v>>16), byte(v>>24)
@@ -33,16 +36,21 @@ func (ch *DMA) set(ofs uint32, b byte) bool {
 	old := byte(ch.cnt() >> 24)
 	ch.io[ofs] = b
 	if ofs == 11 {
-		ch.src, ch.dst = util.LE32(ch.io[0:]), util.LE32(ch.io[4:])
-		switch ch.size() {
-		case 32:
-			ch.src &= ^uint32(3)
-			ch.dst &= ^uint32(3)
-		case 16:
-			ch.src &= ^uint32(1)
-			ch.dst &= ^uint32(1)
+		turnon := !util.Bit(old, 7) && util.Bit(b, 7)
+		if turnon {
+			ch.src, ch.dst = util.LE32(ch.io[0:]), util.LE32(ch.io[4:])
+			ch.count = ch.wordCount()
+			switch ch.size() {
+			case 32:
+				ch.src &= ^uint32(3)
+				ch.dst &= ^uint32(3)
+			case 16:
+				ch.src &= ^uint32(1)
+				ch.dst &= ^uint32(1)
+			}
 		}
-		return !util.Bit(old, 7) && util.Bit(b, 7) && (ch.timing() == 0)
+
+		return turnon && ch.timing() == 0
 	}
 	return false
 }
@@ -81,13 +89,10 @@ func (ch *DMA) timing() dmaTiming { return dmaTiming((ch.cnt() >> (16 + 12)) & 0
 func (ch *DMA) irq() bool         { return util.Bit(ch.cnt(), 16+14) }
 func (ch *DMA) enabled() bool     { return util.Bit(ch.cnt(), 16+15) }
 func (ch *DMA) disable()          { ch.setCnt(ch.cnt() & 0x7fff_ffff) }
-func (ch *DMA) wordCount(i int) int {
+func (ch *DMA) wordCount() int {
 	wordCount := ch.cnt() & 0xffff
 	if wordCount == 0 {
-		wordCount = 0x4000
-		if i == 3 {
-			wordCount = 0x10000
-		}
+		return ch.defaultCount
 	}
 	return int(wordCount)
 }
@@ -98,11 +103,9 @@ func (g *GBA) dmaTransfer(t dmaTiming) {
 			continue
 		}
 
-		g.timer(2)
-
-		wc, size := ch.wordCount(i), ch.size()
+		size := ch.size()
 		srcInc, dstInc := ch.srcCnt(), ch.dstCnt()
-		for wc > 0 {
+		for ch.count > 0 {
 			switch size {
 			case 16:
 				g.setRAM16(ch.dst, g.getRAM16(ch.src, true), true)
@@ -111,20 +114,21 @@ func (g *GBA) dmaTransfer(t dmaTiming) {
 			}
 
 			ch.dst, ch.src = uint32(int64(ch.dst)+dstInc), uint32(int64(ch.src)+srcInc)
-			wc--
+			ch.count--
 		}
 
 		if ch.irq() {
 			g.triggerIRQ(IRQID(irqDMA0 + i))
 		}
 
-		if !ch.repeat() {
+		if ch.repeat() {
+			ch.count = ch.wordCount()
+			if ch.dstReload() {
+				ch.dst = util.LE32(ch.io[4:])
+			}
+		} else {
 			ch.disable()
 		}
-
-		if ch.dstReload() {
-			ch.dst = util.LE32(ch.io[4:])
-		}
 	}
 }
 
diff --git a/pkg/gba/gba.go b/pkg/gba/gba.go
index 435a438..8a340ff 100644
--- a/pkg/gba/gba.go
+++ b/pkg/gba/gba.go
@@ -103,7 +103,7 @@ func (g *GBA) exec(cycles int) {
 	for g.cycle < cycles {
 		g.step()
 		if g.halt {
-			g.cycle = cycles
+			g.timer(cycles - g.cycle)
 		}
 	}
 	g.cycle -= cycles
diff --git a/pkg/gba/io.go b/pkg/gba/io.go
index 39b5648..d073f8e 100644
--- a/pkg/gba/io.go
+++ b/pkg/gba/io.go
@@ -17,13 +17,13 @@ func (g *GBA) _getRAM(addr uint32) uint32 {
 		idx := (bank ^ 0x10) | (addr & 0xf)
 		return util.LE32(waveRAM[idx:])
 	case isDMA0IO(addr):
-		return g.dma[0].get(addr - 0x0400_00b0)
+		return g.dma[0].get(addr - ram.DMA0SAD)
 	case isDMA1IO(addr):
-		return g.dma[1].get(addr - 0x0400_00bc)
+		return g.dma[1].get(addr - ram.DMA1SAD)
 	case isDMA2IO(addr):
-		return g.dma[2].get(addr - 0x0400_00c8)
+		return g.dma[2].get(addr - ram.DMA2SAD)
 	case isDMA3IO(addr):
-		return g.dma[3].get(addr - 0x0400_00d4)
+		return g.dma[3].get(addr - ram.DMA3SAD)
 	case IsTimerIO(addr):
 		return g.timers.GetIO(addr - 0x0400_0100)
 	case addr == ram.KEYINPUT || addr == ram.KEYINPUT+1: