diff --git a/src/gba/gba-io.c b/src/gba/gba-io.c
index 11552f3f5..640c67880 100644
--- a/src/gba/gba-io.c
+++ b/src/gba/gba-io.c
@@ -366,11 +366,23 @@ void GBAIOWrite(struct GBA* gba, uint32_t address, uint16_t value) {
 			GBAIOWrite32(gba, address - 2, gba->memory.io[(address >> 1) - 1] | (value << 16));
 			break;
 
+		// TODO: Confirm this behavior on real hardware
 		case REG_FIFO_A_LO:
 		case REG_FIFO_B_LO:
+			if (gba->performingDMA) {
+				GBAAudioWriteFIFO16(&gba->audio, address, value);
+			} else {
+				GBAIOWrite32(gba, address, (gba->memory.io[(address >> 1) + 1] << 16) | value);
+			}
+			break;
+
 		case REG_FIFO_A_HI:
 		case REG_FIFO_B_HI:
-			GBAAudioWriteFIFO16(&gba->audio, address, value);
+			if (gba->performingDMA) {
+				GBAAudioWriteFIFO16(&gba->audio, address, value);
+			} else {
+				GBAIOWrite32(gba, address - 2, gba->memory.io[(address >> 1) - 1] | (value << 16));
+			}
 			break;
 
 		// DMA
diff --git a/src/gba/gba-memory.c b/src/gba/gba-memory.c
index 1405b9516..82508cd2f 100644
--- a/src/gba/gba-memory.c
+++ b/src/gba/gba-memory.c
@@ -172,7 +172,7 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 
 #define LOAD_BAD \
 	GBALog(gba, GBA_LOG_GAME_ERROR, "Bad memory Load32: 0x%08X", address); \
-	if (cpu->cycles >= cpu->nextEvent) { \
+	if (gba->performingDMA) { \
 		value = gba->bus; \
 	} else { \
 		value = cpu->prefetch[1]; \
@@ -301,7 +301,7 @@ uint32_t GBALoad16(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 			}
 		} else {
 			GBALog(gba, GBA_LOG_GAME_ERROR, "Bad memory Load16: 0x%08X", address);
-			if (cpu->cycles >= cpu->nextEvent) {
+			if (gba->performingDMA) {
 				LOAD_16(value, address & 2, &gba->bus);
 			} else {
 				LOAD_16(value, address & 2, &cpu->prefetch[1]);
@@ -363,7 +363,7 @@ uint32_t GBALoad16(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 		break;
 	default:
 		GBALog(gba, GBA_LOG_GAME_ERROR, "Bad memory Load16: 0x%08X", address);
-		if (cpu->cycles >= cpu->nextEvent) {
+		if (gba->performingDMA) {
 			LOAD_16(value, address & 2, &gba->bus);
 		} else {
 			LOAD_16(value, address & 2, &cpu->prefetch[1]);
@@ -396,7 +396,7 @@ uint32_t GBALoad8(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 			}
 		} else {
 			GBALog(gba, GBA_LOG_GAME_ERROR, "Bad memory Load8: 0x%08x", address);
-			if (cpu->cycles >= cpu->nextEvent) {
+			if (gba->performingDMA) {
 				value = ((uint8_t*) &gba->bus)[address & 3];
 			} else {
 				value = ((uint8_t*) &cpu->prefetch[1])[address & 3];
@@ -460,7 +460,7 @@ uint32_t GBALoad8(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 		break;
 	default:
 		GBALog(gba, GBA_LOG_GAME_ERROR, "Bad memory Load8: 0x%08x", address);
-		if (cpu->cycles >= cpu->nextEvent) {
+		if (gba->performingDMA) {
 			value = ((uint8_t*) &gba->bus)[address & 3];
 		} else {
 			value = ((uint8_t*) &cpu->prefetch[1])[address & 3];
@@ -1138,6 +1138,7 @@ void GBAMemoryServiceDMA(struct GBA* gba, int number, struct GBADMA* info) {
 		}
 	}
 
+	gba->performingDMA = true;
 	int32_t word;
 	if (width == 4) {
 		word = cpu->memory.load32(cpu, source, 0);
@@ -1174,6 +1175,7 @@ void GBAMemoryServiceDMA(struct GBA* gba, int number, struct GBADMA* info) {
 			--wordsRemaining;
 		}
 	}
+	gba->performingDMA = false;
 
 	if (!wordsRemaining) {
 		if (!GBADMARegisterIsRepeat(info->reg) || GBADMARegisterGetTiming(info->reg) == DMA_TIMING_NOW) {
diff --git a/src/gba/gba.c b/src/gba/gba.c
index 1b5f39b7b..aca3af985 100644
--- a/src/gba/gba.c
+++ b/src/gba/gba.c
@@ -76,6 +76,7 @@ static void GBAInit(struct ARMCore* cpu, struct ARMComponent* component) {
 	gba->biosChecksum = GBAChecksum(gba->memory.bios, SIZE_BIOS);
 
 	gba->busyLoop = -1;
+	gba->performingDMA = false;
 }
 
 void GBADestroy(struct GBA* gba) {
diff --git a/src/gba/gba.h b/src/gba/gba.h
index a7b13ed3a..c518fdf8c 100644
--- a/src/gba/gba.h
+++ b/src/gba/gba.h
@@ -112,6 +112,7 @@ struct GBA {
 	struct ARMDebugger* debugger;
 
 	uint32_t bus;
+	bool performingDMA;
 
 	int timersEnabled;
 	struct GBATimer timers[4];