Mackie SP-DSP1 User's Manual

Mackie Industrial White Paper

Noise Sensing Using a Variation of the

nLMS Adaptive Filter with Auto Calibration

CHRIS JUBIEN, AES Member, COSTA LAKOUMENTAS, AES Member,

This White Paper discusses a system that will compensate for the noise level in a room by
measuring the program and noise level sensed through an ambient microphone. The system
then changes the program level in proportion to the noise level that was sensed through the
microphone. The technique that is presented here uses a combination of analog to digital
conversion (ADC), adaptive digital ltering running on a digital signal processor (DSP), and
digital to analog conversion (DAC). The adaptive lter employed is a variation on the Normal­ized Least Mean Squares (nLMS) method. This approach effectively “nulls out” any music
that was sensed at the ambient microphone after which the only thing that remains is the noise.
A Root Mean Square (RMS) measure of this noise level provides the ability to adjust the pro­gram level accordingly.

BRIAN RODEN, AES Member, DALE SHPAK, AES Member,

JEFF SONDERMEYER, AES Member

Mackie Designs, Woodinville, WA

0 Introduction

Virtually anyone who has ever listened to music in an auto­mobile has realized this fundamental fact: while driving, the
music level must be louder than while the car is parked. The
reason for this is because while driving the noise level (wind,
road, etc.) is louder than while parked. This requires that the
listener constantly adjust the music level to compensate for the
varying noise levels. This is not only a problem for automo­biles but virtually all sound systems where background noise
is varying substantially. For example: in a factory setting,
the music would be set to one level while the machinery is
running and another while it is not running. For most systems,
this requires that someone always adjust the level in propor­tion to the noise. The question naturally arises: “why can’t
this be done automatically?” Mackie Designs has invested
a considerable amount of time in research and development
to nd an answer to this very question. In the process, we
have developed a sophisticated DSP noise sensing algorithm
that will perform this task precisely. Mackie’s SP-DSP1™
is an automatic level controller that maximizes intelligibility
by changing gain in proportion to environmental noise level
changes [6]. Basically, this system senses the level of the
ambient noise of a room and adjusts the system gain accord­ingly. To work properly, the controller must “null out” any
effect that the program material (music) has on the noise being

received by the ambient microphone. The method we have
employed to differentiate the noise from the program material
is what makes our algorithm unique (patent pending). One
innovative feature of our algorithm is that it adapts over time
to the varying room acoustics (i.e. people, drapes, sliding
doors, etc.) to provide the best possible music rejection. This
signicantly reduces the possibility of “runaway” gain as
exhibited in existing hardware-based implementations [10].
To accomplish this, we have utilized a combination of digital
hardware (SP-DSP1™) running a complex software algorithm
[8]. Figure 0 shows the hardware block diagram of the noise
sensor. The software is actually twofold: an embedded soft­ware algorithm plus application software (SP-Control™ for
the Palm™) to allow for ease of user control. Additionally,
the SP-DSP1™ algorithm allows for a high level of automation,
which in-turn, makes this system extremely easy to setup
and use. Unlike some of the earlier attempts at “noise sensing”,
the SP-Control™ software requires no complex procedures
during setup and calibration. The user simply places his speaker(s)
as needed and positions the ambient microphone so that it is
listening to the primary noise source. Then the appropriate
gain structure is setup as well as a few room-specic user
parameters. Finally, while playing music, an Auto Calibration
is initiated. It’s fast and simple!

September 2000

1

Mackie Industrial White Paper Noise Sensing

32-bit

Floating-Point

DSP

Stereo

DAC

Stereo

ADC

EPROM

SERIAL

EEPROM

Program Input

Ambient Mic

Input

Program

Output

RX

GND

`

TX

RS-232

TRANS.

RS-232 DSUB9

FIR

Down

Sampler

Down

Sampler

Anti-Alias

Filter

Anti-Alias

Filter

G

RMS

Measure

Coefficient

Calculator

Compander

error

Y

-

+

Noise

Threshold

Override

nLMS Adaptive Filter

Figure 0: Hardware Block Diagram

1 Algorithm Overview

This section summarizes the Ambient Noise Sensing Algorithm
(see Figure 1) implemented in the SP-DSP1™. The algorithm is
used to increase the sensitivity to ambient noise in the room by
rejecting the source signal (music) that is broadcast into the room.
This allows the SP-DSP1™ to control the volume of the music
based on the room ambient noise. The better the rejection of the
music signal the more sensitive the gain control without runaway
gain problems.

The algorithm adapts to the room characteristics by comparing
the room response to the source signal. It computes its own
approximation of the room response in order to cancel the music
signal from the signal picked up by a room microphone [1-5,7].
Room size is the most important factor determining the effective­ness of the algorithm. Small rooms (ofce size) can achieve as
much as 40dB music rejection, while larger rooms (nightclubs
or restaurants) might only achieve 10-20dB. Even 10dB rejection
allows noticeably better sensitivity to noise.

2

Figure 1: Algorithm Block Diagram

September 2000

Mackie Industrial White Paper Noise Sensing

1.1 Algorithm Block Diagram

The ambient noise sensor algorithm is a variation of a Nor­malized LMS Adaptive Filter [2-4, 7, 11-12]. The coefcient
calculator in the system uses the error signal, some Root Mean
Square (RMS) signal levels and energies, and the current coef­cients to compute a new set of Finite Impulse Response (FIR)
coefcients that approach the Room Transfer Function (RTF) [5].
Down sampling has been used to optimize the adaptation char­acteristics of the algorithm for the largest room possible using
minimum MIPS and memory [8, 9]. The compander controls the
gain of the music signal based on the ambient noise detected and
the user input parameters.

1.2 Normalized LMS Adaptive Filter

The Normalized Least Mean Square (nLMS) Adaptive Filter
portion of the Ambient Noise Sensor consists of the FIR lter
(Equation 1), Coefcient Calculator (Equation 2), Summation
(producing the error signal), and RMS measurements, and can
be considered equivalent to the MATLAB1 simulation shown
in Diagram 1 [1, 9, 10]. In this simulation the ‘Voice’ signal
is the ambient noise signal, the delays represent the RTF, and
the nLMS lter adapts to produce the ‘Out’ signal which
is an approximation of the music combined with the RTF.
When the output signal from the nLMS is subtracted from the
input signal (Voice + RTF (Music)) the ‘err’ or error signal
approximates the original ‘Voice’ or noise signal. Look at
the gains and delays shown in the blocks of Diagram 1 (simu­lating the direct sound plus one reection). The amplitude
versus time plot was generated with the coefcients from the
‘Taps’ output of nLMS lter. As you can see from this simple
example, this lter is the RTF. Obviously, in the real world,
the exact RTF is not obtainable. However, the better the
FIR approximates the RTF, the closer the error signal is
to the actual noise in the room [2-4, 11-12]. The nLMS Adap­tive Filter is constantly adapting to the room characteristics.
This provides optimum performance when the room acoustics
change. Room acoustics can change signicantly due to the
arrangement of furnishings, opening or closing drapes, the
number of people in the room changing, or more dramatic
changes such as moveable walls or room dividers.

The effectiveness of the nLMS is determined mostly by the
number of taps in the FIR (internal to the nLMS adaptive
Filter), the adaptation algorithm, and the density and duration
of the RTF. In the SP-DSP1™ the processing capability and
memory allocation has limited the FIR length to about 1500
taps. As a result, the signals are down sampled to allow for
good cancellation even in reasonably large rooms. Since the

1

MATLAB is the trademark of a comercial matrix-based processing

language.

speed of sound is approximately 1 ft/ms, we designed the
system to provide adequate rejection for music paths (direct
sound plus reections) up to about 300 feet.

A standard Transversal n-point FIR lter was used [1, 5, 8].
Input buffering is n points long plus a little extra buffer space.
The Output, y, forms another buffer:

N-1

y(n) = ∑ w

(n)*u(n-i)

i

i=0

(1)
where
N = number of FIR taps or the order of the lter
n = time index
y = FIR output
w = adaptation coefcients
u = FIR input

Calculation of new coefcients (coefcient calculator) is
performed if the following two conditions are met:

• If the normalization energy is greater than zero (if

negative or 0 it is a debug error condition).

• If the Error Threshold signal is above the threshold

computed as a portion of the normalization energy of the
input signal.

Mathematical calculations for coefcient updating are
according to a slightly modied nLMS formula [2-4, 7,
11-12]:

Wi(n+1) = Wi(n)+ ————————

e(n) * beta * u(n-i)

epsilon * E(n)

(2)

where
n = time index
i = 0..n-1
W = FIR lter coefcients
e = error
u = FIR input data
E = Normalization Energy
beta = adaptation rate (function)
epsilon = stability coefcient

Our slightly different use of beta (not shown) plus the use
of an Error Threshold are unique adaptations of the standard
nLMS algorithm. The new beta function allows faster adapta­tion when the error, e, gets smaller. The normal use of beta
as a scaling value (shown above) requires a small value to
guarantee stability of the algorithm when the error is large,

September 2000

3

Mackie Industrial White Paper Noise Sensing

Diagram 1: Matlab simulation Block Diagram and Results

which signicantly reduced the adaptation amount when the
error was small, extending the adaptation time to well over 30
minutes. Our beta function (patent pending) gives stability to
the algorithm when the error is large, but effectively shortens
the adaptation time dramatically when the error becomes
small. This allows the algorithm to adapt to a room in only
a few minutes.

For small rooms a signicant portion of the music energy due
to the paths from speaker to microphone (including reec­tions) are removed by the nLMS algorithm, as there would

4

have to be many reections before the sound could have trav­eled this far. Each reection reduces the energy of the music
signal, and the amount of reduction of the energy is related
to the shape and material of the reecting surface. The more
‘live’ the room, the longer and more dense the RTF is, the
less the energy of the music signal that can be removed by the
nLMS lter [6]. Even without a lot of reections a very large
room may have poor cancellation due to the small number of
reections before the music signal has traveled the maximum
‘distance’ of the FIR.

September 2000