Enhanced energy redistribution speech intelligibility algorithm with real-time implementation

1
Enhanced energy redistribution speech
intelligibility algorithm with real-time
implementation
Manasa Raghavan, Mark D. Skowronski, and John G.
Harris Computational NeuroEngineering Lab,
University of Florida, Gainesville, FL
REAL TIME IMPLEMENTATION The board Texas
Instrument TMS320C6713 DSK Memory usage 30 KB
for program (512 KB total flash memory) 64 KB
for data (8 MB total DRAM) Instructions 225
MHz processor 20 ms window processed in 0.4 ms
(2 or computational bandwidth) Input/output 1/
4 mic input analog signal, microphone or PC 4
dip switches for parameter variation (Gain
terms) 1/4 mic output analog signal, speakers or
headphones
ABSTRACT Speech intelligibility enhancement is a
concern for mobile platforms operating in noisy
environments. Current noise-reduction
techniques, such as subspace methods and spectral
subtraction, operate on speech corrupted by
acoustic and transmission noise. Yet
preprocessing techniques, which operate on clean
speech before noise corruption, have received
little attention. Previously, the authors have
developed the energy redistribution algorithm J.
Acoust. Soc. Am. 112, 2305 (2002), which, based
on characteristics of clear speech as well as the
Lombard effect, redistributes energy in time from
voiced regions to unvoiced regions of speech.
The algorithm is designed efficiently for
real-time implementation, and in this work the
algorithm is demonstrated on a mobile platform,
TIs TMS320C6713 DSK board. Furthermore, two
enhancements to the algorithm are introduced 1)
a variable unvoiced gain factor, and 2) a high
pass filter (HPF). The variable unvoiced gain
factor adjusts the amount of energy
redistributed, and the HPF, a compact algorithm
shown previously to enhance clean speech in noisy
environments Niederjohn and Grotelueschen, IEEE
Trans. Acoust., Speech, Sign. Process., 24 (5),
277-282 (1976), are tested on sentences from the
Hearing in Noise Test (HINT) corrupted by
speech-shaped noise. Results show improved
speech intelligibility for both enhancements.
REAL TIME ENERGY REDISTRIBUTION ALGORITHM
EXPERIMENTS Hearing in Noise Test (HINT) from
House Ear Institute, with audiogram screening. A
custom Matlab test graphical user interface was
created for implementing the HINT test.
Gain and Thresholds The SFM thresholds are
determined from the distribution of SFM values
for the implemented system. Two thresholds allow
for bounce in the SFM signal (see plot
above). The voiced/unvoiced gain terms control
the change in CV ratio. The output signal is
normalized with a final gain term determined by a
running estimate of the output energy. The final
gain term adapts such that the output signal
level is a constant in dB SPL.
DESIGN PARAMETERS 1) Real-time execution --
Low computational cost -- Low latency 2) Apply
knowledge from highly intelligible speech --
Clear speech increased consonant-vowel (CV)
ratio -- Lombard speech decreased spectral
tilt (high pass filter)
REMARKS Questions remain from HINT test
results -- Listener variability significant --
No trend in results from variable Gain factors
(3-20 dB CV boost) -- Combined ERVU/HPF trend
unclear Thresholds for SFM sensitive to input
range, noise, other factors. More principled
adaptive method for demarcating consonants/vowels
needed. Results from previous experiment on
isolated words more promising continuous speech
more difficult to tune system.