diff --git a/sound/usb/card.h b/sound/usb/card.h
index ae4251d5abf7..a39edcc32a93 100644
--- a/sound/usb/card.h
+++ b/sound/usb/card.h
@@ -94,6 +94,8 @@ struct snd_usb_substream {
 	spinlock_t lock;
 
 	struct snd_urb_ops ops;		/* callbacks (must be filled at init) */
+	int last_frame_number;          /* stored frame number */
+	int last_delay;                 /* stored delay */
 };
 
 struct snd_usb_stream {
diff --git a/sound/usb/pcm.c b/sound/usb/pcm.c
index b8dcbf407bbb..0b699ca1957e 100644
--- a/sound/usb/pcm.c
+++ b/sound/usb/pcm.c
@@ -34,6 +34,30 @@
 #include "clock.h"
 #include "power.h"
 
+/* return the estimated delay based on USB frame counters */
+snd_pcm_uframes_t snd_usb_pcm_delay(struct snd_usb_substream *subs,
+				    unsigned int rate)
+{
+	int current_frame_number;
+	int frame_diff;
+	int est_delay;
+
+	current_frame_number = usb_get_current_frame_number(subs->dev);
+	/*
+	 * HCD implementations use different widths, use lower 8 bits.
+	 * The delay will be managed up to 256ms, which is more than
+	 * enough
+	 */
+	frame_diff = (current_frame_number - subs->last_frame_number) & 0xff;
+
+	/* Approximation based on number of samples per USB frame (ms),
+	   some truncation for 44.1 but the estimate is good enough */
+	est_delay =  subs->last_delay - (frame_diff * rate / 1000);
+	if (est_delay < 0)
+		est_delay = 0;
+	return est_delay;
+}
+
 /*
  * return the current pcm pointer.  just based on the hwptr_done value.
  */
@@ -45,6 +69,8 @@ static snd_pcm_uframes_t snd_usb_pcm_pointer(struct snd_pcm_substream *substream
 	subs = (struct snd_usb_substream *)substream->runtime->private_data;
 	spin_lock(&subs->lock);
 	hwptr_done = subs->hwptr_done;
+	substream->runtime->delay = snd_usb_pcm_delay(subs,
+						substream->runtime->rate);
 	spin_unlock(&subs->lock);
 	return hwptr_done / (substream->runtime->frame_bits >> 3);
 }
@@ -417,6 +443,8 @@ static int snd_usb_pcm_prepare(struct snd_pcm_substream *substream)
 	subs->hwptr_done = 0;
 	subs->transfer_done = 0;
 	subs->phase = 0;
+	subs->last_delay = 0;
+	subs->last_frame_number = 0;
 	runtime->delay = 0;
 
 	return snd_usb_substream_prepare(subs, runtime);
diff --git a/sound/usb/pcm.h b/sound/usb/pcm.h
index ed3e283f618d..df7a003682ad 100644
--- a/sound/usb/pcm.h
+++ b/sound/usb/pcm.h
@@ -1,6 +1,9 @@
 #ifndef __USBAUDIO_PCM_H
 #define __USBAUDIO_PCM_H
 
+snd_pcm_uframes_t snd_usb_pcm_delay(struct snd_usb_substream *subs,
+				    unsigned int rate);
+
 void snd_usb_set_pcm_ops(struct snd_pcm *pcm, int stream);
 
 int snd_usb_init_pitch(struct snd_usb_audio *chip, int iface,
diff --git a/sound/usb/urb.c b/sound/usb/urb.c
index e184349aee83..b4dcccc237dc 100644
--- a/sound/usb/urb.c
+++ b/sound/usb/urb.c
@@ -718,7 +718,16 @@ static int prepare_playback_urb(struct snd_usb_substream *subs,
 	subs->hwptr_done += bytes;
 	if (subs->hwptr_done >= runtime->buffer_size * stride)
 		subs->hwptr_done -= runtime->buffer_size * stride;
+
+	/* update delay with exact number of samples queued */
+	runtime->delay = subs->last_delay;
 	runtime->delay += frames;
+	subs->last_delay = runtime->delay;
+
+	/* realign last_frame_number */
+	subs->last_frame_number = usb_get_current_frame_number(subs->dev);
+	subs->last_frame_number &= 0xFF; /* keep 8 LSBs */
+
 	spin_unlock_irqrestore(&subs->lock, flags);
 	urb->transfer_buffer_length = bytes;
 	if (period_elapsed)
@@ -737,12 +746,27 @@ static int retire_playback_urb(struct snd_usb_substream *subs,
 	unsigned long flags;
 	int stride = runtime->frame_bits >> 3;
 	int processed = urb->transfer_buffer_length / stride;
+	int est_delay;
 
 	spin_lock_irqsave(&subs->lock, flags);
-	if (processed > runtime->delay)
-		runtime->delay = 0;
+
+	est_delay = snd_usb_pcm_delay(subs, runtime->rate);
+	/* update delay with exact number of samples played */
+	if (processed > subs->last_delay)
+		subs->last_delay = 0;
 	else
-		runtime->delay -= processed;
+		subs->last_delay -= processed;
+	runtime->delay = subs->last_delay;
+
+	/*
+	 * Report when delay estimate is off by more than 2ms.
+	 * The error should be lower than 2ms since the estimate relies
+	 * on two reads of a counter updated every ms.
+	 */
+	if (abs(est_delay - subs->last_delay) * 1000 > runtime->rate * 2)
+		snd_printk(KERN_DEBUG "delay: estimated %d, actual %d\n",
+			est_delay, subs->last_delay);
+
 	spin_unlock_irqrestore(&subs->lock, flags);
 	return 0;
 }