Speex dev - Sep 2005 - How does the jitter buffer "catch up"?

Thank you for a very good explanation which shed light on some of the
questions that I had after reading the source code.

Reading your text however, I wonder if I'm perhaps missing an important
point on the proper use of the jitter buffer: 

...
> Now, clearly, if early_ratio is high and late_ratio is very 
> low, the buffer is buffering more than it needs to; it will 
> skip a frame to reduce latency.
...

Question:
Do I understand it that I should not put every incoming packet through the
jitter buffer?

The way my code works today is:

1) Packet read from socket
2) Call speex_jitter_put(...) with the just-arrived packet
3) Read one packet from jitter buffer using speex_jitter_get(...) function
4) Feed just read-from-jitter packet to the sound card for playback

This will in fact feed one 20msec batch of sound to play at the sound card
for every packet received from the speex encoder at the other end.

I know I may sound a bit slow-on-the-pickup here, but at the risk of
sounding very beginner like (which I'll gladly admit I am) I wonder if this
is totally wrong to do?

Question:
Should the jitter buffer implementation not have a packet to return (data is
simply missing) should I bother to feed the 20msec packet of silence
(comfort noise perhaps?) to the speaker? Or should the jitter buffer perhaps
hint me (with a return value?) that no packet was available and there is no
need to feed anything to the sound card?

In my current implementation, running on a Windows XP box, I have a growing
number of outstanding packets queued to the soundcard. I believe this is
happening because when packets are delayed (in my test case I have no packet
loss, just delays) the jitter buffer interpolates and returns a packet to
play. When the packets finally arrive, they too are queued to the soundcard.
Resulting in an increasing non-recoverable delay of the speech coming out of
the sound card.

Your feedback is greatly appreciated. I thank you for taking the time to
respond with any relevant details or hints.

Respectfully,
Baldvin
> -----Original Message-----
> From: speex-dev-bounces@xiph.org 
> [mailto:speex-dev-bounces@xiph.org] On Behalf Of Thorvald Natvig
> Sent: 18. september 2005 16:25
> To: speex-dev@xiph.org
> Subject: Re: [Speex-dev] How does the jitter buffer "catch up"?
> 
> 
> > 
> > Is is possible to give a short hint about how the jitter 
> buffer would 
> > "catch up" when network condition have been bad and then get
better?
> >
> > I'm using the jitter buffer with success now, but sometimes 
> I have a 
> > long delay that's caused by bad network conditions and then 
> later when 
> > the conditions get better, I would think we would want the audio to 
> > gradually catch up with real-time to minimize the latency 
> in the voice?
> >
> > Is it not realistic to expect the jitter buffer to do this sort of 
> > "catching up" (of course doing so by "skipping"
some of the older
> > received audio I guess)?
> >
> > I understand the basic idea of the jitter.c code but am 
> aparently not 
> > bright enough to get the whole point of the short- and long-term 
> > margin values etc. Just wonder if it's possible to get a short 
> > description of each of these variables, their purpose and how they 
> > apply to the whole jitter buffer functionality?
> >
> > Thank you very much.
> >
> > Baldvin
> >
> >
> 
> FYI: The below is just my interpretation of the code, I might 
> be wrong.
> 
> Each time a new packet arrives, the jitter buffer calculates 
> how far ahead or behind the "current" timestamp it is; this 
> is called arrival_margin. 
> The  "current" timestamp is simply the last frame 
> successfully decoded.
> 
> It maintains a list of bins for margins, this is short and 
> longterm margin.
> 
> Think of the bins like this:
> 
> -60ms -40ms -20ms 0ms +20ms +40ms +60ms
> 
> when a packet arrives, the margin matching it's 
> arrivel_margin is increased, so if this packet was 40ms after 
> the current timestamp, the 40ms bin would be increased. If 
> this packet arrived 60ms too late (and hence is useless), the 
> -60ms bin would increase.
> 
> early_ratio_XX is the sum of all the positive bins.
> late_ratio_XX is the sum of all the negative bins.
> 
> The difference between _long and _short is just how fast they change.
> 
> If a packet has timestamp outside the bins, it's not used for 
> calculation.
> 
> Now, clearly, if early_ratio is high and late_ratio is very 
> low, the buffer is buffering more than it needs to; it will 
> skip a frame to reduce latency. Alternately, if late_ratio is 
> even marginally above 0, more buffering is needed, and it 
> duplicates a frame. This decision is done when decoding.
> 
> Depending on your chosen transmission method, during network 
> hiccups you'll either have lost packets or they'll come in a 
> burst when the network conditions restore themselves. In 
> either case, after missing 20 packets or so the jitter buffer 
> will prepare to "reset", and it's new current timestamp will 
> be the timestamp on whatever packet arrives. It will also 
> hold decoding until at least buffer_size frames have arrived.
> 
> Since it sounds like you're using reliable transmission 
> (packets are not lost), what will happen is that there's a 
> whole stream of packets suddenly arriving, and they'll fill 
> up the buffer much much faster than it's emptied. In fact, 
> you're likely to fill it so fast the buffer runs out of room, 
> meaning the first few packets gets dropped to make room for 
> the later ones. However, as the current timestamp was set to 
> the first arriving packet, the decoder won't find the packet 
> it's looking for, meaning the jitter buffer will soon reset again.
> 
> So no, it doesn't "catch up", it tries to keep latency to an 
> absolute minimum whatever the circumstances, so most of the 
> late frames will be dropped.
> 
> To achieve the effect you're describing, you'd need to 
> increase SPEEX_JITTER_MAX_BUFFER_SIZE to the longest delay 
> you're expecting, and then inside the block on line 231 (which says)
>     if (late_ratio_short + ontime_ratio_short < .005 && 
> late_ratio_long + ontime_ratio_long < .01 && 
> early_ratio_short > .8) .. add something that multiplies all 
> the magins with 0.75 or so at the end. This will force the 
> jitter buffer to only skip 1 frame at a time and wait a bit 
> before it skips the next one.
> 
> _______________________________________________
> Speex-dev mailing list
> Speex-dev@xiph.org
> http://lists.xiph.org/mailman/listinfo/speex-dev
>

> Thank you for a very good explanation which shed light on some of the
> questions that I had after reading the source code.
>
> Reading your text however, I wonder if I'm perhaps missing an important
> point on the proper use of the jitter buffer:
>
> ...
>> Now, clearly, if early_ratio is high and late_ratio is very
>> low, the buffer is buffering more than it needs to; it will
>> skip a frame to reduce latency.
> ...
>
> Question:
> Do I understand it that I should not put every incoming packet through the
> jitter buffer?
>
> The way my code works today is:
>
> 1) Packet read from socket
> 2) Call speex_jitter_put(...) with the just-arrived packet
> 3) Read one packet from jitter buffer using speex_jitter_get(...) function
> 4) Feed just read-from-jitter packet to the sound card for playback
>
> This will in fact feed one 20msec batch of sound to play at the sound card
> for every packet received from the speex encoder at the other end.
>
> I know I may sound a bit slow-on-the-pickup here, but at the risk of
> sounding very beginner like (which I'll gladly admit I am) I wonder if
this
> is totally wrong to do?
>
> Question:
> Should the jitter buffer implementation not have a packet to return (data
is
> simply missing) should I bother to feed the 20msec packet of silence
> (comfort noise perhaps?) to the speaker? Or should the jitter buffer
perhaps
> hint me (with a return value?) that no packet was available and there is no
> need to feed anything to the sound card?
>
> In my current implementation, running on a Windows XP box, I have a growing
> number of outstanding packets queued to the soundcard. I believe this is
> happening because when packets are delayed (in my test case I have no
packet
> loss, just delays) the jitter buffer interpolates and returns a packet to
> play. When the packets finally arrive, they too are queued to the
soundcard.
> Resulting in an increasing non-recoverable delay of the speech coming out
of
> the sound card.
>
> Your feedback is greatly appreciated. I thank you for taking the time to
> respond with any relevant details or hints.
If you call speex_jitter_put and get from the same thread on every packet, 
you are getting no benefit from the jitter buffer at all.

Correct usage of the jitter buffer is as follows:

Create two threads (or two evens inside the same thread):

Thread/Event #1: On every packet received, immediately and without delay 
call speex_jitter_put.

Thread/Event #2: Every 20ms (use a damn-high-precision timer or 
play-notification from your soundcard), call speex_jitter_get and play 
that frame.

Meaning, even if no packet arrived in the last 20 ms, you should 
still call speex_jitter_get, and no matter if 100 packets arrived in the 
last 20 ms, you should still only call speex_jitter_get once.

(PS, if you do use threads, protect speex_jitter_put/get with a mutex 
(CRITICAL_SECTION I believe they're called in Win32Speak) -- calling put 
and get at the exact same time from different threads leads to
"features")

This allows the jitter buffer to actually buffer, and try to keep the 
delay between frame arrivel and frame play to a minimum.

As for detecting outages, a hack I use is to check jitter->valid_bits. If 
it's set, we decoded "something", if it's not, we're
interpolating
something which may not sound that good so feed the soundcard 20ms of 
silence instead.

Incidentally, on Win32 you probably want to use DirectSound with a looping 
buffer of around 200ms and play notifiers; try to be at least one frame 
ahread of the write cursor -- the notification will come when the play 
cursor is AT the position, meaning that you'd need to decode the packet in 
no time (and I do mean "no time") to avoid artifacts in the sound.
With
one frame to go on you should be safe.

> (PS, if you do use threads, protect speex_jitter_put/get with a mutex 
> (CRITICAL_SECTION I believe they're called in Win32Speak) -- calling
put
> and get at the exact same time from different threads leads to
"features")
I've never tested this, but I designed the jitter buffer to work from
two threads even without using a mutex. This would work as long as there
is one thread doing the "put" and one thread doing the
"get".
> As for detecting outages, a hack I use is to check jitter->valid_bits.
If
> it's set, we decoded "something", if it's not, we're
interpolating
> something which may not sound that good so feed the soundcard 20ms of 
> silence instead.
Why would you do that. The idea of interpolating a frame is exactly to
get better quality than just putting zeros.

	Jean-Marc

-- 
Jean-Marc Valin <Jean-Marc.Valin@USherbrooke.ca>
Universit? de Sherbrooke