manual/include/latency-and-latency-compensation.html

<p>
  <a
  href="http://en.wikipedia.org/wiki/Latency_%28audio%29"><dfn>Latency</dfn></a>
  is a system's reaction time to a given stimulus. There are many factors that
  contribute to the total latency of a system. In order to achieve exact time
  synchronization all sources of latency need to be taken into account and
  compensated for.
</p>

<h2>Sources of Latency</h2>

<h3>Sound propagation through the air</h3>
<p>
  Since sound is a mechanical perturbation in a fluid, it travels at
  comparatively slow <a href="http://en.wikipedia.org/wiki/Speed_of_sound">speed</a>
  of about 340 m/s. As a consequence, an acoustic guitar or piano has a
  latency of about 1&ndash;2&nbsp;ms, due to the propagation time of the sound
  between the instrument and the player's ear.
</p>

<h3>Digital-to-Analog and Analog-to-Digital conversion</h3>
<p>
  Electric signals travel quite fast (on the order of the speed of light),
  so their propagation time is negligible in this context. But the conversions
  between the analog and digital domain take a comparatively long time to perform,
  so their contribution to the total latency may be considerable on
  otherwise very low-latency systems. Conversion delay is usually below 1&nbsp;ms.
</p>

<h3>Digital Signal Processing</h3>
<p>
  Digital processors tend to process audio in chunks, and the size of that chunk
  depends on the needs of the algorithm and performance/cost considerations.
  This is usually the main cause of latency when using a computer and the one that
  can be predicted and optimized.
</p>

<h3>Computer I/O Architecture</h3>
<p>
  A computer is a general purpose processor, not a digital audio processor.
  This means the audio data has to jump a lot of fences in its path from the
  outside to the CPU and back, contending in the process with some other parts
  of the system vying for the same resources (CPU time, bus bandwidth, etc.)
</p>

<h2>The Latency chain</h2>

<p class="note">
  Note! the rest of this document assumes the use of jackd for the audio
  backend. While many of the concepts are true, the specifics may be different.
</p>
<figure>
  <img src="/images/latency-chain.png" alt="Latency chain">
  <figcaption>
    Latency chain
  </figcaption>
</figure>

<p>
  The numbers are an example for a typical PC. With professional gear and an
  optimized system the total round-trip latency is usually lower. The important
  point is that latency is always additive and a sum of many independent factors.
</p>
<p>
  Processing latency is usually divided into <dfn>capture latency</dfn> (the time
  it takes for the digitized audio to be available for digital processing, usually
  one audio period), and <dfn>playback latency</dfn> (the time it takes for the
  audio that has been processed to be available in digital form).
  In practice, the combination of both matters. It is called <dfn>round-trip
  latency</dfn>: the time necessary for a certain audio event to be captured,
  processed and played back.
</p>
<p class="note">
  It is important to note that processing latency in Ardour is a matter of
  choice. It can be lowered within the limits imposed by the hardware (audio
  device, CPU and bus speed) and audio driver. Lower latencies increase the
  load on the system because it needs to process the audio in smaller chunks
  which arrive much more frequently. The lower the latency, the more likely
  the system will fail to meet its processing deadline and the dreaded
  <dfn>xrun</dfn> (short for buffer over- or under-run) will make its
  appearance more often, leaving its merry trail of clicks, pops and crackles.
</p>

<p>
  The digital I/O latency is usually negligible for integrated or
  <abbr title="Periphal Component Interface">PCI</abbr> audio devices, but
  for USB or FireWire interfaces the bus clocking and buffering can add some
  milliseconds.
</p>

<h2>Low Latency use cases</h2>

<p>
  Low latency is <strong>not</strong> always a feature one wants to have. It
  comes with a couple of drawbacks: the most prominent is increased power
  consumption because the CPU needs to process many small chunks of audio data,
  it is constantly active and can not enter power-saving mode (think fan noise).
  Since each application that is part of the signal chain must run in every
  audio cycle, low-latency systems will undergo <dfn>context switches</dfn>
  between applications more often, which incur a significant overhead.
  This results in a much higher system load and an increased chance of xruns.
</p>
<p>
  For a few applications, low latency is critical:
</p>

<h3>Playing virtual instruments</h3>
<p>
  A large delay between the pressing of the keys and the sound the instrument
  produces will throw off the timing of most instrumentalists (save church
  organists, whom we believe to be awesome latency-compensation organic systems.)
</p>

<h3>Software audio monitoring</h3>
<p>
  If a singer is hearing her own voice through two different paths, her head
  bones and headphones, even small latencies can be very disturbing and
  manifest as a tinny, irritating sound.
</p>

<h3>Live effects</h3>
<p>
  Low latency is important when using the computer as an effect rack for
  inline effects such as compression or EQ. For reverbs, slightly higher
  latency might be tolerable, if the direct sound is not routed through the
  computer.
</p>

<h3>Live mixing</h3>
<p>
  Some sound engineers use a computer for mixing live performances.
  Basically that is a combination of the above: monitoring on stage,
  effects processing and EQ.
</p>
<p>
  In many other cases, such as playback, recording, overdubbing, mixing,
  mastering, etc. latency is not important, since it can easily be
  compensated for.
</p>
<p>
  To explain that statement: During mixing or mastering, one doesn&#039;t care
  if it takes 10ms or 100ms between the instant the play button is pressed
  and the sound coming from the speaker. The same is true when recording with a count in.
</p>

<h2>Latency compensation</h2>
<p>
  During tracking it is important that the sound that is currently being
  played back is internally aligned with the sound that is being recorded.
</p>
<p>
  This is where latency compensation comes into play. There are two ways to
  compensate for latency in a DAW, <dfn>read-ahead</dfn> and
  <dfn>write-behind</dfn>. The DAW starts playing a bit early (relative to
  the playhead), so that when the sound arrives at the speakers a short time
  later, it is exactly aligned with the material that is being recorded.
  Since we know that playback has latency, the incoming audio can be delayed
  by the same amount to line things up again.
</p>
<p>
  The second approach is prone to various implementation
  issues regarding timecode and transport synchronization. Ardour uses read-ahead
  to compensate for latency. The time displayed in the Ardour clock corresponds
  to the audio signal that is heard on the speakers (and is not where Ardour
  reads files from disk).
</p>
<p>
  As a side note, this is also one of the reasons why many projects start at
  timecode <samp>01:00:00:00</samp>. When compensating for output latency the
  DAW will need to read data from before the start of the session, so that the
  audio arrives in time at the output when the timecode hits <samp>01:00:00:00</samp>.
  Ardour does handle the case of <samp>00:00:00:00</samp> properly but not all
  systems/software/hardware that you may inter-operate with may behave the same.
</p>

<h2>Latency Compensation And Clock Sync</h2>

<p>
  To achieve sample accurate timecode synchronization, the latency introduced
  by the audio setup needs to be known and compensated for.
</p>
<p>
  In order to compensate for latency, JACK or JACK applications need to know
  exactly how long a certain signal needs to be read-ahead or delayed:
</p>

<figure>
  <img src="/images/jack-latency-excerpt.png" alt="Jack Latency Compensation">
  <figcaption>
    Jack Latency Compensation
  </figcaption>
</figure>

<p>
  In the figure above, clients A and B need to be able to answer the following
  two questions:
</p>
<ul>
  <li>
    How long has it been since the data read from port Ai or Bi arrived at the
    edge of the JACK graph (capture)?
  </li>
  <li>
    How long will it be until the data written to port Ao or Bo arrives at the
    edge of the JACK graph (playback)?
  </li>
</ul>

<p>
  JACK features an <abbr title="Application Programming Interface">API</abbr>
  that allows applications to determine the answers to above questions.
  However JACK can not know about the additional latency that is introduced
  by the computer architecture, operating system and soundcard. These values
  can be specified by the JACK command line parameters <kbd class="input">-I</kbd>
  and <kbd class="input">-O</kbd> and vary from system
  to system but are constant on each. On a general purpose computer system
  the only way to accurately learn about the total (additional) latency is to
  measure it.
</p>

<h2>Calibrating JACK Latency</h2>
<p>
  Linux DSP guru Fons Adriaensen wrote a tool called <dfn>jack_delay</dfn>
  to accurately measure the round-trip latency of a closed loop audio chain,
  with sub-sample accuracy. JACK itself includes a variant of this tool
  called <dfn>jack_iodelay</dfn>.
</p>
<p>
  Jack_iodelay allows to measure the total latency of the system,
  subtracts the known latency of JACK itself and suggests values for
  jackd's audio-backend parameters.
</p>
<p>
  jack_[io]delay works by emitting some rather annoying tones, capturing
  them again after a round trip through the whole chain, and measuring the
  difference in phase so it can estimate with great accuracy the time taken.
</p>
<p>
  The loop can be closed in a number of ways:
</p>
<ul>
  <li>
    Putting a speaker close to a microphone. This is rarely done, as air
    propagation latency is well known so there is no need to measure it.
  </li>
  <li>
    Connecting the output of the audio interface to its input using a
    patch cable. This can be an analog or a digital loop, depending on
    the nature of the input/output used. A digital loop will not factor
    in the <abbr title="Analog to Digital, Digital to Analog">AD/DA</abbr>
    converter latency.
  </li>
</ul>
<p>
  Once the loop has been closed, one must:
</p>
<ol>
  <li>Launch jackd with the configuration to test.</li>
  <li>Launch <kbd class="input">jack_delay</kbd> on the command line.</li>
  <li>Make the appropriate connections between the jack ports so the loop is closed.</li>
  <li>Adjust the playback and capture levels in the mixer.</li>
</ol>
<p class="warning">
  On Linux, the latency of USB audio interfaces is not constant. It may
  change when the interface is reconnected, on reboot and even when xruns
  occur. This is due the buffer handling in the Linux USB stack. As a
  workaround, it is possible to recalibrate the latency at the start of each
  session and each time an xrun occurs.
</p>