Upload
scot-ball
View
220
Download
1
Tags:
Embed Size (px)
Citation preview
Dual-domain Hierarchical Classification of Phonetic
Time Series
Hossein Hamooni, Abdullah Mueen University of New Mexico
Department of Computer Science
What is Phoneme? Phonemes are very small units of intelligible sound (usually less than 200 ms).
Phonetic spelling is the sequence of phonemes that a word comprises.
Example: Coat ([kōt] /K OW T/) From ([frəm] /F R AH M/) impressive ([imˈpresiv] /IH M P R EH S IH V/)
2
Phoneme Classification
What is phoneme classification?
Input: A short segment of audio signal.
Output: What phoneme it is.
Phoneme classification is a complex task:
More than 100 classes (based on International Phonetic Alphabet)
Variation in speakers, dialects, accents, noise in the environment, etc.
Phoneme classification can be used in:
Robust speech recognition
Accent/dialect detection
Speech quality scoring
3
Related Work
Different methods for phoneme classification have been used in the literature: Hidden Markov model [Lee, 1989]
Neural network [Schwarz, 2009]
Deep belief network [Mohamed, 2012]
Support vector machine [Salomon, 2001]
Hierarchical methods [Dekel, 2005]
Boltzmann machine [Mohamed, 2010]
Although data mining society has shown that k-NN classifiers can work well on time series data, it hasn’t been tried on phoneme yet.
4
[C. Lopes, F. Perdigao, 2011]
Our Dual-domain Approach
5
Time Domain: Using k-NN Dynamic Time Warping (DTW) Expensive Speed up by lower bounding
techniques
Frequency Domain: Using k-NN Euclidean distance between Mel-
frequency cepstrum coefficients (MFCC)
Fast
Real Example
6
Challenge
7
DTW is expensive (quadratic in time and space complexity)
We need to apply a speed up technique Solution: Lower bounding techniques
w w
DTW Lower bounding
8
Resampling to equal length doesn’t always work !!!
DTW Lower bounding
9
We use the prefix of the longer signal (Prefixed LB_Keogh) We show that Prefixed LB_Keogh is a lower bound if:
w > difference between lengths of two signals We set w = c * length of the longer signal We ignore all pairs of signals that don’t satisfy the above condition.
2 4 6 8 10 12 14 16 18x104
0
0.5
1
1.5
2
2.5
3
3.5
Sp
eed
up
Training Set Size10 20 30 40 50 60 70 80 90 100
80.2
80.4
80.6
80.8
81
81.2
81.4
81.6
81.8
Window Size (c%)
Acc
urac
y(%
)
c = 30%
Data Collection
10
370,000 phonemes are segmented from: Data is publicly available.
AH T S IH IY M EH AE AA FOW V AO
UW W HH CHAW OY ZH
05000
1000015000200002500030000350004000045000
Num
ber o
f sam
ples
Phoneme Segmentation
11
The Penn Phonetics Lab Forced Aligner (p2fa) Takes a signal and a transcript Produces timing segmentations (word level and phoneme level)
Accuracy (All layers)
12
10-fold cross validation 100 random phonemes in each fold
Accented Phoneme Classification
13
0 0.5 1 1.5 2 2.5 3 3.5x 104
0.65
0.7
0.75
0.8
0.85
0.9
0.95
Training Set Size
Acc
urac
y
MFCC
DTW
British vs. American accent Using Oxford test set 2-class classification problem No hierarchy
Conclusion We present a dual-domain hierarchical method for phoneme
classification.
We generate a novel dataset of 370,000 phonemes.
We achieve up to 73% accuracy rate for 39 classes.
Our lower bounding technique gives us up to 3X speedup.
14
15
Thank You
Data and code available at:http://cs.unm.edu/~hamooni/papers/
Dual_2014