Department of Electrical Engineering | University of Texas

Department of Electrical Engineering | University of Texas

Department of Electrical Engineering | University of Texas at Dallas
Erik Jonsson School of Engineering & Computer Science | Richardson, Texas 75083-0688,
U.S.A.

Design and fabricate a new microphone array system which will:
Minimize the feature size of the array system by integration of the
microphone and preamplifier circuit.
Reduce noise factors and electro magnetic interferences
Be a portable microphone array and power regulation system
Determine the array configuration and array processing method
that will give the optimum speech to noise ratio.
Implement a speaker recognition system to determine who is speaking
among a closed set of known drivers.
Explore the use of Wireless Transmission, VoIP, and Packet Loss
Concealment

Microphone &
Preamplifier Circuit
Bandpass Filter for Human Voice
(20Hz 20kHz)

Differential Amplifier
Gain: 45 V/V

Data Acquisition
(DEWTRON)

Power Regulator

CSA-BF showed significant improvement over DASB in the segmental signal-to-noise ratio test and
the logarithmic array setup showed a small increase in the segmental SNR over the linear array setup.

Given a corrupted signal, we used two
different PLC Algorithms to fill the gaps.

In the future, audio collected during UTDrive may be sent wirelessly for collection rather than recorded to a hard drive
physically present in the vehicle.
VoIP allows transmission of speech over the Internet real-time, which can be effective for recording to a hard drive remotely.
VoIP audio, however, suffers quality loss when packets are dropped from poor connections. The received audio has gaps
where the packets are dropped and causes the audio to have significantly degraded quality.
As quality continues to drop, it becomes harder to understand what the speaker is saying. A further loss in quality makes it
difficult to understand who is even speaking, let alone what they are saying. To help understand these premises we:
Created a survey to determine a minimum quality of voice that a listener was willing to listen to and could understand.
Surveyed listeners to see what minimum quality is needed to at least understand who the speaker is, given an unaffected
voice sample.
Surveyed listeners to see which packet loss concealment scheme (a simple one) is useful for recovering lost quality.
Packet loss concealment is used to make up for times when packets are dropped---trying to fill in the "blanks" can
potentially recover lost quality.
Speaker recognition is a process by which the identity of a speaker can be determined. This project is only concerned with
closed set speaker recognition, meaning the speaker is assumed to belong to a known set of people whose voices have
already been collected and processed to produce Gaussian Mixture Models (GMMs). The speaker recognition software will
perform feature extraction on short voice samples and compare them to the existing GMMs to determine a best fit.

Speaker recognition test with UTDrive corpus
database using 2-8 seconds of speech

Because the close talk microphone sample was not very clean, the comparison of the beam formed sample to
channel 3 resulted in a higher segmented SNR value than that of the beam formed sample and the close talk
microphone. This problem with the close talk microphone will be addressed in the next phase of the project.
HTK was used to perform speaker recognition on clean out of car data. Results for in car recordings could not
be obtained for various reasons, including engine noise, interference, and competing speakers in the vehicle.
These problems will be addressed next semester with the new microphone array and beam forming software.

Recently Viewed Presentations