USB Synchronous Multichannel Audio

Acquisition System
David Abran-Côté, Massinissa Bandou, Alexandre Béland, Gabriel Cayer, Sébastien Choquette,
Frédéric Gosselin, Francis Robitaille, Diallo Telly Kizito, François Grondin, Dominic Létourneau

University of Sherbrooke

Abstract — This paper is a technical description of a XMOS XS1- eight preamplifier microphone cards, all powered by the main
L2 based synchronous multichannel audio acquisition system audio card.
designed for mobile applications, or more specifically mobile
robots. The design is made available through an Open Source The USB 2.0 High-Speed interface is used for data transfer as
Hardware license, allowing easy modifications and sharing. The it is more commonly available compared to other interfaces
system includes an 8 channels USB sound card coupled with 8
such as FireWire or Peripheral Component Interconnect (PCI)
active and differential microphones. The sound card also
provides an analog stereo output. It can be powered by either and more portable. The USB 2.0 transfer rate reaches 480
USB or an external wide range (7-36V) DC power supply. The Mbits/sec, which is sufficient to transfer the raw
electrical consumption over USB does not exceed 2.5 W, (uncompressed) data of 8 microphones and one stereo output.
conforming to the USB 2.0 power specifications enabling the
system to be solely powered by the USB bus. Differential codecs This paper presents a description of the designed system. The
and amplification circuitry are used, allowing operation in noisy project specifications, global architecture, materials and
setups. Finally the sound card is compatible with the following software, tests results and analysis are presented.
operating systems: Windows, Linux and Mac OS. Audio data is
transferred to the computer using the USB 2.0 High-Speed Audio B. State of the Art
Class 2.0 standard avoiding designing a different driver for each Robot artificial audition systems use studio sound card since
operating system. there is no audio acquisition card with eight input channels
specifically designed for robotic applications. These
Keywords – Synchronous data acquisition, artificial audition, professional cards often provide unnecessary functionalities
audio codecs, sound sources localization, open source, mobile (sound effects, integrated mixing, optical inputs/outputs,
robotics, differential signals, USB Audio Class 2.0, USB 2.0 High
Speed, XMOS
S/PDIF, MIDI, numerous analogs outputs, etc.). Moreover,
they are large, expensive, and require a significant amount of
I. CONTEXT power. Clearly, these devices do not meet the requirements for
the current application.
A. Introduction and Motivation
Previously, sound cards like the RME MULTIFACE II (PCI)
ntRoLab [1] is a research laboratory based in
I Sherbrooke, Québec, Canada. IntRoLab is pursuing the
goal of studying, developing and integrating
and the MOTU Ultra-Lite-mk3 Hybrid (FireWire) were used
at IntRoLab with the ManyEars system. These cards cost
around US $500-$1000, consume too much power (10W+,
technologies for the design of autonomous and intelligent without the power consumption of the eight microphones),
systems. Research activities involve software and hardware provide inadequate or complicated connectivity (power supply
design of mobile robots, embedded systems and autonomous + PC interface) and dimensions are not suitable for robotic
agents. To enhance the robot perception, an artificial audition applications.
system named ManyEars [2] has been developed to locate,
track and separate sound sources in real time with an array of So far, the USB synchronous audio acquisition interfaces use
eight microphones. This capability allows the robot to locate a mostly either the USB Audio Class 1.0 standard, which is
person or an interesting event in its environment. Tracking the limited to two input channels, or a proprietary USB protocol.
sound sources over time also lets the robot follow different A few other systems use the USB Audio Class 2.0 standard
speakers in motion. ManyEars is also able to separate the but these are not well suited for this application.
individual speech of multiple simultaneous speakers. Tests
show that the system performs well with up to four sound The definition of the audio class norm is applicable to all USB
sources located within seven meters. computer compatible devices that have integrated functions
for voice or audio manipulation.
Since the algorithm uses an array of eight microphones, an
eight channel input synchronous audio acquisition system is Recently, the USB Audio Class 2.0 standard was introduced in
needed. Two major challenges are present in mobile robotics: few audio acquisition devices [3]. The improvements over the
1) power consumption must be low and 2) physical Audio Class 1.0 standard include more channels and better
dimensions of the system must be optimized. The goal of this sampling resolutions and frequencies. Most mobile robots use
project is thus to design a small, low-cost and low-power a personal computer (PC) as their central processing system.
audio acquisition card with microphones. To accommodate the Linux is the most popular operating system as it exclusively
needs of the ManyEars algorithm, the card has eight input supports the Robotic Operating System (ROS) [4].
channels and one stereo output channel and is connected to
 Figure 1: System diagram

Until now, a few manufacturers provide devices using USB

C. Components
Audio Class 2.0 but interest is growing with the development
1) USB Communication
of systems with multiples channels that needs more resolution
The card has to communicate by USB using the USB 2.0
and flexibility.
protocol. The audio input and output channels are driven by
the USB Audio Class standard, which supports eight input
II. DESCRIPTION OF THE SYSTEM
channels since version 2.0 [3].
A. Request for Proposal
2) Microcontroller
The main objective is to design a USB synchronous audio The system uses a XMOS microprocessor with second
acquisition card, which has eight input channels and one generation « L » architecture [5]. A reference design is
analog stereo output channel, and eight pre-amplification available and meets the needs of the project in terms of
microphone cards. The acquisition card needs to be powered hardware and software. This reference design also includes a
by the USB port or an external source. The microphones have software implementation of the Audio Class 2.0 standard,
to be powered by the main USB card. The total consumption which is not available with other processor manufacturer than
of the system must not exceed 5W when powered by the USB XMOS.
interface in overcharge (5V at 1A). The card dimensions
should be approximately 7cm x 10cm x 3cm. Signal resolution The XMOS architecture is promising as it is based on events
should be a minimum of 12 effective bits and the sampling [6] instead of processor cycles like most microcontrollers [7].
frequency should range from 48 to 192 kHz. Event programming improves power management as the
B. Global Hardware Architecture processor dynamically manages power according to the
frequency of the events. On the other hand, this architecture
The hardware architecture, shown at Figure 1, is divided in limits the number of available threads and their execution
three main components: the USB communication, the times are fixed by the hardware. Consequently, the processor
microcontroller and the audio codecs. These components are
might not be used at its full potential because each defined
shown in Figure 1.
thread has the same capacity regardless of the quantity of
processed data. The programmer must then take the XMOS
architecture into account to optimize power consumption and
performances.
The assembly of the double core processor involves additional analog signal with a resolution of 16 bits, we have 66 µV per
costs due to the QFN-124 physical format, which requires division [11]. The 4.3V LDO (TPS79501) is also chosen for
specialized equipment. Moreover, the assembly implies a its small size, good immunity to input power supply noise and
validation test using X-rays. rejection ratio (PSRR).

3) Audio acquisition A digital reference voltage of 3.3V is also required to power

Differential signals are used to reduce noise. The codec the codecs, the USB PHY, the phase locked loop (PLL) and
CS42448 from Cirrus Logic [8] has differential inputs, the input/output ports of the XMOS processor. The conversion
provides tools to filter noise and supports both the I2S to 3.3V is achieved by a step-down DC-DC converter
protocol [9] for audio transfer and the I2C protocol [9] for the (NCP1421E).
codec configuration. Furthermore, experiments with the
XMOS audio 2.0 development kit confirmed that the Cirrus Two other converters are required. A LDO converter
Logic Codec and XMOS processor are compatible. Some (NCP599SN18) is used to produce a voltage of 1.8V to power
analog filters proposed by Cirrus Logic are added to the the USB PHY and a DC-DC converter (FAN2011) is used to
differential inputs to reduce crosstalk in the microphone power the XMOS cores.
cables.
2) Microphone card
As mentioned earlier, the sound card is designed to acquire
audio signals from eight microphones placed at fixed known
locations. Each microphone needs its own preamplifier circuit
board, which is powered by the main acquisition card. The
physical dimension of these cards needs to be minimized.
Figure 2 shows a picture of one 2.3cm x 2.3cm microphone
card. The major electronic components include the
microphone (of electret type), the preamplifier (TS472) and
the RJ11 connector (on the back side). The preamplifier is
selected since it has a high signal-to-noise ratio, differential
input and output channels and a closed loop gain of
approximately 40dB (to obtain a peak-to-peak amplitude of
4.3V at the output). The RJ11 connector is proposed since
Figure 2: Differential input buffer parallel insertion prevents the power line from making contact
with the data line. The connector should also have a latch
D. Hardware mechanism. The RJ11 and XLR mini both meet these
1) Power Supply requirements and RJ11 is selected due to its low cost and easy
The synchronous multichannel audio card is designed to be use.
powered either by USB or an external power supply. The first
scenario simplifies the integration as the existing circuitry and
wiring are used. However, an external power supply improves
the controllability of the power output and avoids overloading
the USB hub. Powering the device from USB is limited to
2.5W and yields to stricter power specifications as the
maximum allowed current is 500 mA and the voltage
randomly ranges from 4.4V to 5.25V.

The voltage of an external power supply ranges from 7V to

36V, which allows different input voltages provided by the
robot batteries. To minimize the voltage and current ripples, Figure 3: Microphone card
the input and output signals of the regulation circuit are
filtered. 3) Analog Signal Conditioning and Integrity

When the external power supply is turned on, a power Strategies to preserve the audio signal integrity are described
multiplexer (TPS2111) automatically detects this new source this section.
and disables the USB power source. This allows hot plugging
of power sources. The analog signal coming from the microphones is transmitted
in differential mode to the codecs. Differential mode is
To maximize the signal-to-noise ratio and to obtain the 12 preferred to single-ended signaling because of its better noise
effective bits resolution previously specified, a high analog immunity and because it increases the dynamic range of the
reference voltage is needed. This is done with a low-dropout analog/digital converter of the codec. Since the acquisition
voltage regulator (LDO) that ensures that the analog reference card is designed to operate on a robot, many external devices
voltage stays at 4.3V when the USB voltage is at its minimum can induce electromagnetic interference in the transmitted
voltage (4.4V). Consequently, when the ADCs acquire the
signal. Twisted pairs for signal transmission distribute the divided into two sections, one for the digital parts and another
interference and a differential amplifier rejects the common one for analog parts.
mode noise, which minimizes the effect of overall
electromagnetic interference. The power plane is divided in three sections, as shown in
figure 3. The first section is used for the digital parts (in red),
Moreover, the amplitude of a differential signal is higher than the second one is used for the analog parts (in orange) and the
the amplitude of a single-ended signal for the same reference third one powers the XMOS cores (in blue).
voltage, which increases the dynamic range. According to the
datasheet of the codec, the dynamic range goes from 102dB
(single-ended signaling) to 105dB (differential signaling),
which yields to a 41% increase.

The microphone preamplifiers are positioned as close as

possible to the electret in order to reduce the effects of
electromagnetic interference and to preserve a good signal-to-
noise ratio.

At the preamplifier’s output, the signal goes through a

differential audio amplifier (LME49726) that biases the signal
for the codec’s input and also filters the signal to avoid
aliasing. This amplifier is chosen because it is especially
intended for audio and it uses a single power supply. The
band-pass filter has a flat frequency response in the audio
Figure 4: Power Supply Plane
band and has a rejection of 20dB at the lowest sampling
frequency. The configuration of the anti-aliasing filters is the The digital section of the power plane provides 3.3V to the
one suggested by the manufacturer. digital parts of the two codecs, the USB PHY, the phase
locked loop (PLL) and the XMOS digital I/O. This section
4) Codecs
covers approximately half the PCB.
Cirrus Logic CS42448 codecs are used to convert the signals
The analog section of the power plane provides 4.3V to the
from the analog domain to the digital domain in both
analog parts of the codecs, the differential amplifiers, and the
directions and the I2S (Inter IC Sound) serial protocol is used
two audio outputs.
to exchange data with the XMOS processor. To meet the
specifications, each codec uses four out of the six 24 bits The XMOS section of the power plane provides 1V to the
converter input channels. The sampling frequency is XMOS cores and is located directly under the integrated
configurable and can go up to 192 kHz. circuit of the XMOS processor for an easier access to inner
pins.
The various codec settings (converters’ operation mode,
sampling frequency, power and more) are set by the XMOS This division improves the signal-to-noise ratio (SNR) since
processor through a serial communication I2C (Inter the high frequency noise generated by the digital parts does
Integrated Circuit) link. not interfere to the analog components.
The codec needs to operate with an analog reference voltage A constant length between each differential pair is ensured to
of 4.3V (as opposed to 5V for other codecs). It has a dynamic minimize the signal distortion. Decoupling capacitors are also
range of 105 dB in differential mode. Moreover, the codec is used to ensure proper operation of the card.
part of a reference design from XMOS Company. Codecs are
preferred over ADC with eight channels because the card 1) Software Components
needs to have an audio output channel, which would require 1) USB Audio Class 2.0
an additional DAC. However, the current design somewhat
underuse the codecs, because each codec could support up to 6 The prototype uses Audio Class 2.0 of the USB 2.0 standard
analog inputs and 6 analog outputs. A future version of the and is already compatible with Linux (Ubuntu 10.10 and up)
sound card could then support up to 12 inputs and 12 outputs. and Mac OS. It is not supported by Windows (Win7 SP1) but
We prioritized size over number of inputs / outputs or this a driver for this operating system is provided by a third party.
iteration of the project. Hoping that Microsoft provides support soon due to the rising
popularity of Audio Class 2.0.
5) PCB (Printed Circuit Board) Design
The XMOS reference design provides Audio Class 2.0. This
The PCB has four layers: a ground plane, a power plane and implementation is available under a particular license that
two planes for routing. A single ground plane is preferred over prevents us from releasing the source code. However, it’s
separated ground planes to avoid ground loops. The PCB is possible to take a confidential agreement with XMOS to get
access to the source code (NDA). Nevertheless, this determines the loss in effective bits in the least significant bit
implementation is used with the current design. (LSB) zone. The desired effective number must be higher than
the 12 bits operating threshold for ManyEars[15]. A fixed-
The generic driver available with Audio Class 2.0 is used. frequency signal produced by a frequency generator is used as
There is thus no need to design a custom driver. This option the analog input signal of the system. A slight increase of the
also reduces the amount of support to be provided after the signal amplitude is done. The smallest amplitude variation
project. (voltage) that changes the digital signal by one bit is
measured. The range of voltage (4.33V) is then divided by this
2) Event Based Architecture voltage variation in order to obtain the effective number of
The « L » architecture of XMOS microprocessor family is of bits (ENOB).
particular interest for mobile applications. The eight threads
for each core are independently activated by events instead 4) Mean Noise Floor
and thus do not run continuously with the system clock. This To quantify the mean noise floor of the system, an analog
reduces power consumption since only active threads require input channel of the card receives a signal from the generator
power. An external clock is used to synchronize the codecs and another input channel acts as the control signal. The
and XMOS cores for the data acquisition [12-13]. difference between the two inputs is used to calculate the
mean noise floor of the system.
III. RESULTS, VALIDATION AND TESTS B. Quantitative Results
The first test consists in measuring the voltage of each
The following table displays the results of these tests
power line via test points.
performed on the acquisition system.
The external power supply is then unplugged when the USB Size Length (mm) Width Height
interface is connected. This ensures the power multiplexor (mm) (mm)
works properly and that the device supports hot plugging. Sound card 125 74 15
Microphone 23 23 15
Sound is then played in a loudspeaker and recorded with the
Power supply Voltage (V) Current (A) Consumption
audio card connected via USB to a PC. The recorded signals External 7 0.332 2.324 W
are then listened back. (without 12 0.198 2.376 W
microphones) 24 0.114 2.736 W
All tests are passed, which confirms that the system meets the 36 0.0867 3.121 W
project requirements. The audio acquisition of the eight analog External (with 7 0.343 2.401 W
microphones) 12 0.208 2.496 W
inputs and the analog stereo output are functional as well as 24 0.118 2.843 W
the audio data transfer from the USB port. The system is 36 0.0884 3.182 W
functional on Linux, Mac and Windows. USB 5 0.434 2.1675 W
(without
A. Validation Method and Selected Tests microphones)
This section presents the key tests performed to characterize USB 5 0.449 2.246 W
(with
the USB synchronous multichannel audio acquisition system. microphones)
Maximum with external power supply 3.182 W
1) Electrical Consumption Maximum with USB power supply 2.246 W
The electrical consumption is verified for the external power
supply and USB when the system is plugged with and without Latency Maximum delay between 2 audio channels
16 µs
the eight microphones. The external power supply can range
from 7 to 36 V. The consumption is thus tested for different ENOB Effective number of bits
external voltages (7, 12, 24, and 36 V) and for USB power 22 Effective bits
only. The total consumption of the system must not exceed
5W. Noise Floor
Mean -132 dBV
Maximum -112 dBV
2) Latency between Acquisition Channels
Figure 5: Results table
The same signal is transmitted to the eight microphone inputs
installed at the same distance from an omnidirectional sound The card’s dimensions are slightly larger than the ones
source. Recording is performed on all channels with the requested in the initial specifications (i.e. 125mm × 74mm vs
Electroacoustic Toolbox [14] software. The integrated 100mm × 70mm). However, this small oversize has a
oscilloscope is used to confirm that the offset between the negligible impact on the quality of the final product. We have
inputs is not higher than one acquisition cycle. chosen bigger and easy to use RJ-12 connectors that could be
replaced by something smaller to gain PCB space easily.
3) Effective Number of Bits (ENOB)
The ratio between the maximal noise in the input signals of the On the other hand, the maximum power consumption of the
audio codec versus the reference voltage of the same codec sound card is less than the one in the original specifications
(i.e. approximately 3.2W against 5W). Maximum power [3] USB. (2006, Mai 31). Audio Devices Rev. 2.0 Spec and Adopters
Agreement [Online]. Available :
consumption occurs when using the maximal external voltage
http://www.usb.org/developers/devclass_docs/Audio2.0_final.zip
of 36V. If the external voltage is reduced to the minimal [4] Robot Operating System (ROS), [Online], Available :
voltage of 7V, consumption drops to 2.5W. The variation of http://www.ros.org/wiki/
consumption is caused by the reduction of power loss in the [5] XMOS. (2011, Mars 17). Description of the “L” architecture. [Online].
Available : http://www.xmos.com/technology/architecture
DC-DC converter at the input of the power supply.
[6] XMOS. (2011, Mars 17). Event based core. [Online]. Available :
http://www.xmos.com/technology/xcore
When the system is powered through USB, the maximum [7] XMOS. (2011, Mars 17). XMOS vs FPGA. [Online]. Available :
consumption is approximately 2.3W. This implies that the https://www.xmos.com/download/public/XMOS-vs-FPGA-
sound card can be powered without using USB in overcharge Whitepaper%281%29.pdf
[8] Cirrus Logic. (2007, November). CS43448. [Online]. Available :
(a current of 500 mA or less is sufficient). http://www.cirrus.com/en/pubs/proDatasheet/CS42448_F3.pdf
[9] NXP. (1996, June 5). I2S Specifications. [Online]. Available :
There is a maximum delay of 16µs between two acquisition http://www.nxp.com/acrobat_download2/various/I2SBUS.pdf
channels. Thus, for a sampling frequency of 96kHz, the [10] NXP. (2000, January). I2C Specifications. [Online]. Available :
maximum delay between two channels is higher than one http://www.nxp.com/documents/other/39340011.pdf
[11] Clayton R. Paul, « Electromagnetics for Engineers » Wiley, 2004.
sample. However, for a target sampling frequency of 48kHz, [12] XMOS (2011, January 14). USB Audio Software Design Guide.
the time delay is smaller than one sample. This test was done [Online]. Available : https://www.xmos.com/published/usb-audio-
with a common signal for two distinct inputs on each physical software-design-guide
codec. However, we could have pushed the test further by [13] XMOS. (2010, June 29). USB Audio 2.0 MC Ref Design Schematics.
[Online]. Available : https://www.xmos.com/published/usb-audio-20-
testing delay between each of the 8 inputs simultaneously. mc-ref-design-schematics
[14] Electroacoustic Toolbox 3.0. [Online]. Available :
The initial goal of 12 effective bits is met as the resolution http://www.faberacoustical.com/products/electroacoustics_toolbox/
goes up to 22 bits. The mean value of the noise floor obtained [15] Texas Instrument. (2009, Mai). The Basics of the Effective Number of
Bits. [Online]. Available :
during the tests is similar to the one specified by the codec. http://e2e.ti.com/videos/m/analog/97246.aspx?DCMP=hpa_dc_general
This means that the card has good noise rejection. &HQS=NotApplicable+OT+bitselection-pr
[16] XMOS. (2011, Mai 7). XS1-L2 124QFN Datasheet. [Online]. Available:
IV. CONCLUSIONS AND FUTURE WORK https://www.xmos.com/published/xs1-l2-124qfn-datasheet
A. Conclusion
The USB synchronous multichannel audio acquisition system
meets the objectives on the project specifications. The system
allows the acquisition of eight analog input channels
synchronously at the frequency of 48kHz. The system
provides a resolution of 22 effective bits which exceeds easily
the specification of 12 bits required by the ManyEars
algorithm. The final sound card has a size slightly larger than
the targeted size.

The system can be powered by an external or USB power

supply. The maximal power consumption with USB is 2.3W,
which meets the requirements.

Finally, the system is fully operational and it can be used with

the ManyEars system by the IntRoLab mobile robotics and
intelligent systems laboratory.
B. Future Development
This project could lead to future developments with the
XLINK ports [16]. These ports allow the addition of
microcontrollers in a serial topology and thus can increase the
number of parallel operations in the system. Moreover, the
current threads do not use all the resources on the processor.
These resources could be used to perform some computations
for sound source localization directly on the acquisition card.

REFERENCES
[1] IntRoLab. (2009, August 26). [Online]. Available :
http://www.introlab.gel.usherbrooke.ca
[2] ManyEars. (2010, November 19). [Online]. Available :
http://www.manyears.sourceforge.net