diff --git a/components/squeezelite/output.c b/components/squeezelite/output.c
index d54dcef0..08e9665e 100644
--- a/components/squeezelite/output.c
+++ b/components/squeezelite/output.c
@@ -254,10 +254,7 @@ frames_t _output_frames(frames_t avail) {
 		
 		out_frames = !silence ? min(size, cont_frames) : size;
 
-		if (output.channels & 0x01) gainR |= MONO_FLAG;
-		if (output.channels & 0x02) gainL |= MONO_FLAG;
-
-		wrote = output.write_cb(out_frames, silence, gainL, gainR, cross_gain_in, cross_gain_out, &cross_ptr);
+		wrote = output.write_cb(out_frames, silence, gainL, gainR, output.channels, cross_gain_in, cross_gain_out, &cross_ptr);
 
 		if (wrote <= 0) {
 			frames -= size;
diff --git a/components/squeezelite/output_bt.c b/components/squeezelite/output_bt.c
index eed7ba88..50d81611 100644
--- a/components/squeezelite/output_bt.c
+++ b/components/squeezelite/output_bt.c
@@ -40,7 +40,7 @@ static uint8_t *btout;
 static frames_t oframes;
 static bool stats;
 
-static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 								s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr);
 								
 #define DECLARE_ALL_MIN_MAX \
@@ -79,7 +79,7 @@ void output_close_bt(void) {
 	equalizer_close();
 }	
 
-static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 						 s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr) {
 	
 	assert(btout != NULL);
@@ -90,7 +90,7 @@ static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t g
 			_apply_cross(outputbuf, out_frames, cross_gain_in, cross_gain_out, cross_ptr);
 		}
 
-		_apply_gain(outputbuf, out_frames, gainL, gainR);
+		_apply_gain(outputbuf, out_frames, gainL, gainR, flags);
 
 #if BYTES_PER_FRAME == 4
 		memcpy(btout + oframes * BYTES_PER_FRAME, outputbuf->readp, out_frames * BYTES_PER_FRAME);
@@ -112,7 +112,7 @@ static int _write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t g
 		memcpy(btout + oframes * BYTES_PER_FRAME, buf, out_frames * BYTES_PER_FRAME);
 	}
 	
-	output_visu_export(btout + oframes * BYTES_PER_FRAME, out_frames, output.current_sample_rate, silence, ((gainL & ~MONO_FLAG) + (gainR & ~MONO_FLAG)) / 2);
+	output_visu_export(btout + oframes * BYTES_PER_FRAME, out_frames, output.current_sample_rate, silence, (gainL  + gainR) / 2);
 	
 	oframes += out_frames;
 
diff --git a/components/squeezelite/output_i2s.c b/components/squeezelite/output_i2s.c
index cb51984c..49dc2c72 100644
--- a/components/squeezelite/output_i2s.c
+++ b/components/squeezelite/output_i2s.c
@@ -101,7 +101,7 @@ static struct {
 
 DECLARE_ALL_MIN_MAX;
 
-static int _i2s_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _i2s_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 								s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr);
 static void *output_thread_i2s(void *arg);
 static void output_thread_i2s_stats(void *arg);
@@ -398,20 +398,20 @@ bool output_volume_i2s(unsigned left, unsigned right) {
 /****************************************************************************************
  * Write frames to the output buffer
  */
-static int _i2s_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR,
+static int _i2s_write_frames(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags,
 								s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr) {
 	if (!silence) {
 		if (output.fade == FADE_ACTIVE && output.fade_dir == FADE_CROSS && *cross_ptr) {
 			_apply_cross(outputbuf, out_frames, cross_gain_in, cross_gain_out, cross_ptr);
 		}
 		
-		_apply_gain(outputbuf, out_frames, gainL, gainR);
+		_apply_gain(outputbuf, out_frames, gainL, gainR, flags);
 		memcpy(obuf + oframes * BYTES_PER_FRAME, outputbuf->readp, out_frames * BYTES_PER_FRAME);
 	} else {
 		memcpy(obuf + oframes * BYTES_PER_FRAME, silencebuf, out_frames * BYTES_PER_FRAME);
 	}
 
-	output_visu_export(obuf + oframes * BYTES_PER_FRAME, out_frames, output.current_sample_rate, silence, ((gainL & ~MONO_FLAG) + (gainR & ~MONO_FLAG)) / 2);
+	output_visu_export(obuf + oframes * BYTES_PER_FRAME, out_frames, output.current_sample_rate, silence, (gainL + gainR) / 2);
 	oframes += out_frames;
 	
 	return out_frames;
diff --git a/components/squeezelite/output_pack.c b/components/squeezelite/output_pack.c
index c84d7c4e..dcf48056 100644
--- a/components/squeezelite/output_pack.c
+++ b/components/squeezelite/output_pack.c
@@ -43,34 +43,31 @@ s32_t to_gain(float f) {
 	return (s32_t)(f * 65536.0F);
 }
 
-void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, output_format format) {
+void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, u8_t flags, output_format format) {
 	// in-place copy input samples if mono/combined is used (never happens with DSD active)
-	if ((gainR & MONO_FLAG) && (gainL & MONO_FLAG)) {
+	if ((flags & MONO_LEFT) && (flags & MONO_RIGHT)) {
 		s32_t *ptr = inputptr;
 		frames_t count = cnt;
-		gainL &= ~MONO_FLAG; gainR &= ~MONO_FLAG;
 		while (count--) {
 			// use 64 bits integer for purists but should really not care
 			*ptr = *(ptr + 1) = ((s64_t) *ptr + (s64_t) *(ptr + 1)) / 2;
 			ptr += 2;
 		}
-	} else if (gainL & MONO_FLAG) {
+	} else if (flags & MONO_RIGHT) {
 		s32_t *ptr = inputptr + 1;
 		frames_t count = cnt;
-		gainL &= ~MONO_FLAG;
 		while (count--) {
 			*(ptr - 1) = *ptr;
 			ptr += 2;
 		}
-	} else if (gainR & MONO_FLAG) {
+	} else if (flags & MONO_LEFT) {	
 		s32_t *ptr = inputptr;
 		frames_t count = cnt;
-		gainR &= ~MONO_FLAG;
 		while (count--) {
 			*(ptr + 1) = *ptr;
 			ptr += 2;
 		}
-    }
+	}	
 	
 	switch(format) {
 #if DSD
@@ -383,37 +380,37 @@ void _apply_cross(struct buffer *outputbuf, frames_t out_frames, s32_t cross_gai
 #if !WIN
 inline 
 #endif
-void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR) {
-	if (gainL == FIXED_ONE && gainR == FIXED_ONE) {
+void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR, u8_t flags) {
+	if (gainL == FIXED_ONE && gainR == FIXED_ONE && !(flags & (MONO_LEFT | MONO_RIGHT))) {
 		return;
-	} if ((gainR & MONO_FLAG) && (gainL & MONO_FLAG)) {
+	} else if ((flags & MONO_LEFT) && (flags & MONO_RIGHT)) {
 		ISAMPLE_T *ptrL = (ISAMPLE_T *)(void *)outputbuf->readp;
 		ISAMPLE_T *ptrR = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
-		gainL &= ~MONO_FLAG; gainR &= ~MONO_FLAG;
 		while (count--) {
 			*ptrL = *ptrR = (gain(gainL, *ptrL) + gain(gainR, *ptrR)) / 2;
 			ptrL += 2; ptrR += 2;
 		}
-	} else if (gainL & MONO_FLAG) {
+
+	} else if (flags & MONO_RIGHT) {
 		ISAMPLE_T *ptr = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
 		while (count--) {
 			*(ptr - 1) = *ptr = gain(gainR, *ptr);
 			ptr += 2;
 		}
-	} else if (gainR & MONO_FLAG) {
+	} else if (flags & MONO_LEFT) {
 		ISAMPLE_T *ptr = (ISAMPLE_T *)(void *)outputbuf->readp;
 		while (count--) {
 			*(ptr + 1) = *ptr = gain(gainL, *ptr);
 			ptr += 2;
 		}
-   } else {
-   		ISAMPLE_T *ptrL = (ISAMPLE_T *)(void *)outputbuf->readp;
+	} else {
+	   	ISAMPLE_T *ptrL = (ISAMPLE_T *)(void *)outputbuf->readp;
 		ISAMPLE_T *ptrR = (ISAMPLE_T *)(void *)outputbuf->readp + 1;
 		while (count--) {
 			*ptrL = gain(gainL, *ptrL);
 			*ptrR = gain(gainR, *ptrR);
 			ptrL += 2; ptrR += 2;
 		}
-   }	
+	}
 }
 
diff --git a/components/squeezelite/squeezelite.h b/components/squeezelite/squeezelite.h
index ddb7a4b3..b34d6451 100644
--- a/components/squeezelite/squeezelite.h
+++ b/components/squeezelite/squeezelite.h
@@ -472,7 +472,6 @@ void _wake_create(event_event*);
 #define MAX_SILENCE_FRAMES 2048
 
 #define FIXED_ONE 	0x10000
-#define MONO_FLAG	0x20000
 
 #ifndef BYTES_PER_FRAME
 #define BYTES_PER_FRAME 8
@@ -655,6 +654,8 @@ typedef enum { FADE_INACTIVE = 0, FADE_DUE, FADE_ACTIVE } fade_state;
 typedef enum { FADE_UP = 1, FADE_DOWN, FADE_CROSS } fade_dir;
 typedef enum { FADE_NONE = 0, FADE_CROSSFADE, FADE_IN, FADE_OUT, FADE_INOUT } fade_mode;
 
+#define MONO_RIGHT	0x02
+#define MONO_LEFT	0x01
 #define MAX_SUPPORTED_SAMPLERATES 18
 #define TEST_RATES = { 768000, 705600, 384000, 352800, 192000, 176400, 96000, 88200, 48000, 44100, 32000, 24000, 22500, 16000, 12000, 11025, 8000, 0 }
 
@@ -675,7 +676,7 @@ struct outputstate {
 	unsigned latency;
 	int pa_hostapi_option;
 #endif
-	int (* write_cb)(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr);
+	int (* write_cb)(frames_t out_frames, bool silence, s32_t gainL, s32_t gainR, u8_t flags, s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr);
 	unsigned start_frames;
 	unsigned frames_played;
 	unsigned frames_played_dmp;// frames played at the point delay is measured
@@ -758,9 +759,9 @@ void output_close_stdout(void);
 #endif
 
 // output_pack.c
-void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, output_format format);
+void _scale_and_pack_frames(void *outputptr, s32_t *inputptr, frames_t cnt, s32_t gainL, s32_t gainR, u8_t flags, output_format format);
 void _apply_cross(struct buffer *outputbuf, frames_t out_frames, s32_t cross_gain_in, s32_t cross_gain_out, ISAMPLE_T **cross_ptr);
-void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR);
+void _apply_gain(struct buffer *outputbuf, frames_t count, s32_t gainL, s32_t gainR, u8_t flags);
 s32_t gain(s32_t gain, s32_t sample);
 s32_t to_gain(float f);