AEC & Q-SYS Conferencing System (Part 2)

Welcome back. Now that we’ve looked at the process in a little detail,

let’s look at some standard conferencing system terminology:

Let’s say that I’m in a room using a Q-SYS system for conferencing.

I am what’s called the ‘near end’.

This means my AEC algorithm will cancel echo for the callers into my room.

Note that the AEC used in my system doesn’t help me, it’s there only for the remote participants.

The ‘far end’ is the remote party in my conference.

Again, the AEC in my system as the ‘near end’ is to benefit them.

As noted before, the AEC algorithm doesn’t know what signal it should be looking for and attempting to cancel.

We do this by connecting that signal to the AEC reference pin of the processing block.

We’ll discuss best practice for this later in the training module.

The process of ‘convergence’ is what the adaptive filter is doing

to achieve the best echo cancellation possible given the circumstances.

When the far end is hearing no echo, we say the AEC algorithm is converged.

As we’ve discussed, proper AEC processing is an intensive operation.

In general as DSP processes get more intensive the audio latency,

or the time it takes a signal to travel through a signal path increases.

Some of this latency is purposeful in order to give the algorithm more time between the arrival

of the reference signal and the microphone signal that needs to be processed.

Note that the input to ouptut audio latency through Q-SYS is 3.17ms without AEC in the signal path.

The AEC algorithm introduces 21.4ms into the

signal path of the 110f and 10.7ms for all other Q-SYS cores.

That’s a total latency of 24.57ms in the core 110f and 13.87ms in all other cores.

That may seem like a long time,

but remember conferencing systems aren’t typically latency sensitive applications.

It usually takes far longer for the signal to travel to the far end,

so this is just one component of the overall conferencing latency.

A larger room will usually have reflections that return to the microphone later than a small one.

The tail length of the AEC algorithm determines how long it ‘looks’ for reflections to be cancelled.

The Q-SYS AEC algorithm has adjustable tail lengths of 100, 200, 300 and 400ms.

Remember the number of AEC algorithms a given core will support is specified for the 200ms tail length.

300 and 400ms algorithms require more processing resources, so fewer microphones can be processed.

For example, the core 110f can support 16 200ms algorithms

while the number decreases to 8 for 400ms tail lengths.

The Q-SYS AEC algorithm employs a concept known as ‘Reference to Microphone Level Ratio’,

or RMLR to help us properly calibrate the system.

This is analogous to another term common to AEC systems – ‘Echo Return Loss’ or ERL.

It represents the natural loss through the output gain structure, through the room and back to the microphone.

The RMLR represents the difference between the level at the reference

and the resulting signal heard at the microphone.

In this example, the level of the reference is -20db while the resultant level at the microphone is -30db.

The difference, or RMLR in this case is 10dB.

To get the best performance from the algorithm, our goal is to calibrate the block to achieve an RMLR of 0.

Later we’ll see how to make this adjustment.

The Echo Return Loss Enhancement, or ERLE, represents

the relative success of the algorithm in cancelling the echo.

Mathematically, it is the difference between the level of the signal at the input of the block

and the resultant level.

In this case, we see a level of -30dB into the algorithm with a level of -50dB post processing.

The computed ERLE would then be 20dB.

We can think of is this way: The RMLR represents

the natural relationship between the level in the room vs. the level of the echo.

The ERLE represents the work done by the algorithm itself.

We saw before that the AEC algorithm isn’t always completely successful. ‘Residual Echo Suppression’,

or RES, is the non linear processing that removes that left over signal.

It’s called ‘non-linear’ because this part of the algorithm breaks the residual signal

into many small frequency bands and adjusts the level of each independently.

For example, there may be -40dB of signal at 500Hz while there’s -50dB of signal at 2kHz.

RES will attenuate the 500Hz band more than 2k.

As we continue to explore AEC applications in this module,

we are considering rooms small enough that participants

can hear presenters without the need for local reinforcement.

There are generally two conferencing system applications that require in-room reinforcement:

Voice lift is the term often used when we’re referring to

local sound reinforcement of a presenter in a given room.

This is not exclusive to conferencing systems,

but is still quite common in rooms with a large presenter’s area

that’s far from the conferencing participants.

In very large conference rooms or other applications such as legislature spaces,

there are enough participants that not all of them can hear each other without some reinforcement.

In most cases a scheme known as ‘mix minus’ is employed so that a given talker

is only reinforced in parts of the room where they aren’t heard directly.

This requires a sophisticated mix and many discrete speaker circuits to get the best results.

Best practice and troubleshooting for these two scenarios are covered in the Quantum level 2 curriculum.

So, stay tuned for news on that.

For now let's just take a quick break, and come back whenever you're ready.