Powerpoint template for scientific posters (Swarthmore College)

1
The Robustness of MFCCs in Phoneme-Based Speaker
Recognition using TIMIT
Rio Akasaka 09, Youngmoo Kim, Ph.D Department
of Linguistics/Engineering, Swarthmore College
Drexel University
Results
Conclusions While optimal performance in speaker
recognition is expected with a larger training
set, the availability of testing material did not
seem to affect performance if at least three
files are used and if the number of training
files is equal to or greater than the number of
testing files. Though this might be expected to
extend to the length of the wav files, it was not
necessarily the case because using half a file to
test consistently demonstrated poor
results. Most importantly, testing and training
with vowel phones only provided impressive
recognition rates at approximately 93, meriting
further study. With regards to individual phone
contributions to recognition, it was found that a
single phoneme does not predict a speaker more
effectively when using the same phoneme to train,
as compared to any other phoneme. However, two
phonemes consistently outperform the others in
predicting 1a speaker 'ae' and 'ay'. Of the five
trials, 'ae' was ranked most highly recognized 3
times, 'ay' was highest twice, and both were
among the top two in four of the trials. More
tests are being done to obtain a statistically
significant conclusion.

Introduction Mel- Frequency Cepstral Coefficients
(MFCCs) are quantitative representations of
speech and are commonly used to label sound
files. They are derived by obtaining the Fourier
transform of a signal and mapping the result on
the mel-scale, which is an auditory
perception-based scale of pitch differences. With
these unique labels on speech files, the
similarity between two files can be determined by
the KullbackLeibler (KL) distance, which is
based on probability distributions, and, given a
training set upon which to base ones decisions,
the corresponding speaker can be identified. The
goal of this research is to test the robustness
of MFCCs in speaker detection by varying the
testing and training parameters with the
following methods 1) using segments of a whole
speech file 2) varying the number of speech
files used, and 3) splicing together the vowel
phones of a speech segment
Vowels only, cont. If individual phoneme files
are used for training and testing, the results
are impressive. Train 5V, Test 5V Evaluated
570 Correct 533, Percentage 0.935088 Train 3V,
Test 3V Evaluated 342 Correct 278, Percentage
0.812865
The following nomenclature is adopted in this
poster F Full (complete) speech file H Speech
file segmented at middle V File consisting of
vowel phones only
Control Train 5F , Test 5F (not including SA)
Evaluated 570 Correct 493, Percentage
0.864912 Difference in number Reducing both the
number of training and testing files to be
consistent results in an optimal (84.2) success
rate, but only up to 3 files. Train 3F , Test
5F Evaluated 570 Correct 440, Percentage
0.771930 Train 5F , Test 3F Evaluated 342
Correct 292, Percentage 0.853801 Train 3F ,
Test 3F Evaluated 342 Correct 290, Percentage
0.847953
Further examination In order to extract more
information about the role that individual phones
play in speaker recognition, the same algorithm
was applied to test recognition based on
individual phonemes that are extracted from each
speaker. The training set consists of files
containing only file segments for a particular
phoneme, which are then later tested
individually.
The TIMIT Corpus The TIMIT corpus was created as
a joint effort between Texas Instruments (TI) and
MIT and consists of time-aligned orthographic,
phonetic and word transcriptions for each of the
6300 16-bit 16kHz speech files. 630 speakers from
the 8 major dialects of American English each
read from 10 phonetically rich texts, among
which 2 are common across all speakers.
PHONETIC 10160 10733 y 10733 11880
axr WORD 10160 11880 your ORTHOGRAPHIC 0 57140
She had your dark suit in greasy wash water
all year. In order to investigate the
distribution of the phonemes in TIMIT, the plot
shown above was generated. The average sample
length is The individual texts may be
phonetically rich, but taken as a whole the
distribution of the phonemes is unbalanced.
Literature cited Cole, Ronald A., et al.. 1996.
The Contribution of Consonants Versus Vowels to
Word Recognition in Fluent Speech Van Heerden,
C.J, E. Bernard. 2008. Speaker-specific
variability of Phoneme Durations. Fattah,
Mohamed, Ren Fuji, Shingo Kuroiwa. 2006. Phoneme
Based Speaker Modeling to Improve Speaker
Identification
Figure 2. Speaker prediction based on individual
phonemes. The results show that while speaker
recognition based on individual phoneme is
considerably low (µ3.60, s2.34), the diagonal
does show slightly higher recognition rates, as
would be expected.
Figure 1. Speaker recognition based on 144
vowel-based files
Figure 1. The predicted speaker ID plotted
against the actual speaker, for 144 full speech
files.
Difference in size Performance is considerably
better with full speech files during testing,
regardless of which half of the file we
use. Train 5F, Test 5H Evaluated 570 Correct
327, Percentage 0.573684 Train 5H, Test
5F Evaluated 570 Correct 442, Percentage
0.775439 Train 5H, Test 3F Evaluated 342
Correct 277, Percentage 0.809942
Acknowledgments Grateful acknowledgement is made
to Youngmoo Kim for providing insight and
direction throughout my research and to Jiahong
Yuan for encouraging my pursuit of corpus
phonetics.
For further information
Vowels only Individual phonemes are exceedingly
difficult to use when predicting speaker based on
entire wav files. Train 5V, Test 3F Evaluated
342 Correct 242, Percentage 0.707602 Train
5F, Test 5V Evaluated 570 Correct 53,
Percentage 0.092982
Figure 3. To test the possibility that one
speaker is consistently retrieved as the ideal
candidate for a particular phoneme, the above
plot was generated to plot the predicted speaker
vs the actual speaker based on speaker ID.
Speaker 183 is selected most often in the above
scenario.
Please contact rakasak1_at_swarthmore.edu. Further
details about the methodology may be read online
at wiki.rioleo.org