An Efficient Online Algorithm for Hierarchical Phoneme Classification

1
An Efficient Online Algorithm for Hierarchical
Phoneme Classification

• Joseph Keshet
• joint work with Ofer Dekel and Yoram Singer
• The Hebrew University, Israel

MLMI 04 Martigny, Switzerland
2
Motivation
Phonetic transcription of DECEMBER
Gross errors
d ix CH eh m bcl b er
Minor errors
d AE s eh m bcl b er
d ix s eh NASAL bcl b er
3
Hierarchical Classification

• Goal spoken phoneme recognition

PHONEMES
Sononorants
Silences
Nasals
Obstruents
Liquids
n
m
ng
l
Vowels
y
w
Affricates
r
Plosives
jh
Fricatives
ch
Front
Center
Back
f
b
v
g
sh
oy
aa
iy
d
s
ow
ao
ih
k
th
uh
er
ey
p
dh
uw
aw
eh
t
zh
ay
ae
z
4
Metric Over Phonetic Tree

• A given hierarchy induces a metric over the set
  of phonemes ? tree distance

5
Metric Over Phonetic Tree

• A given hierarchy induces a metric over the set
  of phonemes ? tree distance

6
Metric Over Phonemes

• Metric semantics?(a,b) is the severity of
  predicting phoneme group b instead of correct
  phoneme a

• Our high-level goal
• Tolerate minor errors
• Sibling errors
• Under-confident predictions - predicting a parent
• but, avoid major errors

7
Hierarchical Classifier

• Assume and
• Associate a prototypewith each phoneme
• Score of phonemeas
• Classification rule

8
Hierarchical Classifier

• Goal maintain close to
• Define
• Goal maintain small

9
Online Learning

• For
• Receive an acoustic vector
• Predict a phoneme
• Receive correct phoneme
• Suffer tree-based penalty
• Apply update rule to obtain

Goal Suffer a small cumulative tree error
10
Tree Loss

• Difficult to minimize
  directly
• Instead upper bound by
  wherealso known as the hinge loss

11
Online Update
w0
w1
w2
w6
w7
w4
w5
w8
w3
w10
w9
12
Loss Bound Theorem

• sequence of examples
• satisfies
• Then
• where and

13
Extension Kernels

• Since
• Note that
• Therefore

14
Experiments

• Synthetic data
• Symmetric tree of depth 4, fan out 3, 121 labels
• Prototypes orthogonal set in with
  Gaussian noise
• 100 train instances and 50 test instances per
  label
• Phoneme recognition
• Subset of the TIMIT corpus
• 55 phonemes and phoneme groups
• MFCC??? front-end, concatenation of 5 frames
• RBF kernel
• 2000 train vectors and 500 test vector per phoneme

15
Experiments

• Multiclass - Ignore the hierarchy

C
16
Results
17
Results
Difference between the tree error rates of the
tree algorithm and the multiclass (MC) algorithm
gross errors
Tree err-MC err
Tree err-MC err
minor errors
Synthetic data
Phonemes
18
Tree vs. Multiclass Online Learning

• Similarity between the prototypes in Multiclass
  and Tree training

19
Thanks!