Designing IP-Based Video Conferencing Systems: Dealing with Lip Synchronization（唇音同步）

转自：http://www.ciscopress.com/articles/article.asp?p=705533&seqNum=6

Correlating Timebases Using RTCP

The RTCP protocol specifies the use of RTCP packets to provide information that allows the sender to map the RTP domain of each stream into a common reference timebase on the sender, called the Network Time Protocol (NTP) time. NTP time is also referred to aswall clock time because it is the common timebase used for all media transmitted by a sending endpoint. NTP is just a clock measured in seconds.

RTCP uses a separate wall clock because the sender may synchronize any combination of media streams, and therefore it might be inconvenient to favor any one stream as the reference timebase. For instance, a sender might transmit three video streams, all of which must be synchronized, but with no accompanying audio stream. In practice, most video conferencing endpoints send a single audio and video stream and often reuse the audio sample clock to derive the NTP wall clock. However, this generalized discussion assumes that the wall clock is separate from the capture clocks.

NTP

The wall clock, which provides the master reference for the streams on the sender endpoint, is in units of NTP time. However, it is important to bear in mind what NTP time isand what NTP time is not:

NTP time as defined in the RTP specification is nothing more than a data format consisting of a 64-bit double word: The top 32 bits represent seconds, and the bottom 32 bits represent fractions of a second. The NTP time stamp can therefore represent time values to an accuracy of ± 0.1 nanoseconds (ns).
The most widespread misconception related to the RTCP protocol is that it requires the use of an NTP time server to generate the NTP clock of the sender. An NTP time server provides a service over the network that allows clients to synchronize their clocks to the time server. The time server specifies that NTP time should measure the number of seconds that have elapsed since January 1, 1970. However, NTP time as defined in the RTP spec does not require the use of an NTP time server. It is possible for RTP implementations to use an NTP time server to provide a reference timebase, but this usage is not necessary and is out of scope of the RTP specification. Indeed, most video conferencing implementations do not use an NTP time server as the source of the NTP wall clock.

Forming RTCP Packets

Each RTP stream has an associated RTCP packet stream, and the sender transmits an RTCP packet once every few seconds, according to a formula given in RFC 3550. As a result, RTCP packets consume a small amount of bandwidth compared to the RTP media stream.

For each RTP stream, the sender issues RTCP packets at regular intervals, and those packets contain a pair of time stamps: an NTP time stamp, and the corresponding RTP time stamp associated with that RTP stream. This pair of time stamps communicates the relationship between the NTP time and RTP time for each media stream. The sender calculates the relationship between its NTP timebase and the RTP media stream by observing the value of the RTP media capture clock and the NTP wall clock in real time. The clocks have both an offset and a scale relationship, according to the following equation:

RTP/(RTP sample rate) = (NTP + offset) x scale

After determining this relationship by calculating the offset and scale values, the sender creates the RTCP packet in two steps:

The sender first selects an NTP time stamp for the RTCP packet. The sender must calculate this time stamp carefully, because the time stamp must correspond to the real-time value of the NTP clock when the RTCP packet appears on the network. In other words, the sender must predict the precise time at which the RTCP packet will appear on the network and then use the corresponding NTP clock time as the value that will appear inside the RTCP packet. To perform this calculation, the sender must anticipate the network interface delay.
After the sender determines the NTP time stamp for the RTCP packet, the sender calculates the corresponding RTP time stamp from the preceding relationship as follows:

RTP = ((NTP + offset) x scale) x sample_rate

The sender can now transmit the RTCP packet with the proper NTP and RTP time stamps.

Determining the values of offset and scale is nontrivial because the sender must figure out the NTP and RTP time stamps at the moment the capture sensor (microphone or camera) captures the data. For instance, to determine the exact point in time when the capture device samples the audio, the sender might need to take into account delays in the capture hardware. Typically, the audio capture device makes a new packet of audio data available to the main processor and then triggers an interrupt to allow the processor to retrieve the packet. When the sender processes an interrupt, the sender must calculate the NTP time of the first sample in each audio packet, corresponding to the moment in time when the sample entered the microphone. One method of calculating this time is by observing the time of the NTP wall clock and then subtracting the predicted latency through the audio capture hardware. However, a better way to map the captured samples to NTP time is for the capture device to provide two features:

A way for the sender to read the device clock of the capture device in real time, and therefore correlate the capture device clock to NTP wall clock time.
A way for the sender to correlate samples in the captured data to the capture device clock. The capture device can provide this functionality by adding its own capture device time stamp to each chunk of audio data.

From these two features, the sender can correlate audio samples to NTP wall clock time. The sender can then establish the relationship between NTP time and RTP time stamps by assigning RTP time stamps to the data.

The same principles apply to the video capture device. The sender must correlate a frame of video to the NTP time at which the camera CCD imager captures each field. The sender establishes the RTP/NTP mapping for the video stream by assigning RTP values to the video frames.

NOTE

The Microsoft DirectX streaming technology used for capture devices defines source filters, which are capture drivers that generate packets of captured data, along with time stamps. Hardware vendors write source filters for their capture hardware. Applications that use these source filters rely entirely on the source filters to provide data with accurate time stamps. If a source filter provides output streams for audio and video, it is critical that the source filter use kernel-level routines to ensure that the time stamps on the packets accurately reflect the time at which the hardware samples the media.

Using RTCP for Media Synchronization

The method of synchronizing audio and video is to consider the audio stream the master and to delay the video as necessary to achieve lip sync. However, this scheme has one wrinkle: If video arrives later than audio, the audio stream, not the video stream, must be delayed. In this case, audio is still considered the master; however, the receiver must first add latency to the audio jitter buffer to make the audio "the most delayed stream" and to ensure that synchronization can be achieved by delaying video, not audio.

In addition, the receiver must determine a relationship between the local audio device timebase ATB and the local video device timebase VTB on the receiver by calculating an offset AtoV:

VTB = ATB/(audio sample rate) + AtoV

This equation converts the local audio device timebase ATB into units of seconds by dividing the audio device time stamp by the audio sample rate. The receiver determines the offset AtoV by simultaneously observing Vtime, the value of the real-time video device clock, and Atime, the value of the real-time audio device clock. Then

AtoV = Vtime – ATime/(audio sample rate)

Now that the receiver knows AtoV, it can establish the final mapping for synchronization.

NOTE

The Microsoft DirectX streaming technology used for playout devices defines render filters, which are essentially playback drivers that accept packets of data with time stamps that are relative to a global system time. The render filters play the media at the time indicated on the time stamp. Hardware vendors write filters for their playout hardware. Applications that use these render filters rely entirely on the render filters to play data accurately, based on the time stamps. A DirectX streaming render filter provides input connections in the form of input pins. If a render filter provides input pins for audio and video, it is critical that the render filter use kernel-level procedures to ensure that the time at which the hardware displays the media is accurately reflected by the time stamp on the packet.

To establish this mapping, two criteria must be met:

At least one RTP packet must arrive from each stream.
The receiver must receive at least one RTCP packet for each stream, to associate each RTP timebase with the common NTP timebase of the sender.

For this method, the audio is the master stream, and the video is the slave stream. The general approach is for the receiver to maintain buffer-level management for the audio stream and to adapt the playout of the video stream by transforming the video RTP time stamp to a video device time stamp that properly slaves to the audio stream.

When a video frame arrives at the receiver with an RTP time stamp RTPv, the receiver maps the RTP time stamp RTPv to the video device time stamp VTB using four steps, as illustrated in Figure 7-14.

Figure 7-14 Audio and Video Synchronization

This sequence of steps maps the RTP video time stamp into the audio RTP timebase and then back into the video device timebase. The receiver follows these steps in order:

Map the video RTP time stamp RTPv into the sender NTP time domain, using the mapping established by the RTP/NTP time stamp pairs in the video RTCP packets.
From this NTP time stamp, calculate the corresponding audio RTP time stamp from the sender using the mapping established by the RTP/NTP time stamp pairs in the audio RTCP packets. At this point, the video RTP time stamp is mapped into the audio RTP timebase.
From this audio RTP time stamp, calculate the corresponding time stamp in the audio device timebase by using the Krl offset. The result is a time stamp in the audio device timebase ATB.
From ATB, calculate the corresponding time stamp in the video device timebase VTB using the offset AtoV.

The receiver now ensures that the video frame with RTP time stamp RTPv will play on the video presentation device at the calculated local video device timebase VTB.

Lip Sync Policy

The receiver may decide not to attempt to achieve lip sync for synchronized audio and video streams in certain circumstances, even if lip sync is possible. There are two scenarios in which this situation might occur:

Excessive audio delay—If the receiver must delay audio to establish lip sync, the receiver might instead choose to achieve the lower audio latency of unsynchronized streams. The reason is because lower end-to-end audio latency achieves the best real-time interaction. The receiver can make this determination after it achieves buffer management for both audio and video streams. If the audio stream is the most-delayed stream, the receiver can opt to delay the video stream to achieve lip sync; if the video stream is the most-delayed stream, however, the receiver might opt to avoid delaying audio to achieve lip sync.
Excessive video delay—If the receiver must delay video by a significant duration to achieve lip sync, on the order of a second or more, the receiver might need to store a large amount of video bitstream in a delay buffer. For high bit rate video streams, the amount of memory required to store this video data might exceed the available memory in the receiver. In this case, the receiver may opt to set an upper limit on the maximum delay of the video stream to accommodate the limited memory or forego video delay altogether.