Korean Phoneme Discrimination

1
Korean Phoneme Discrimination

• Ben Lickly

2
Motivation

• Certain Korean phonemes are very difficult for
  English speakers to distinguish
• ?(IPA s)
• ?(IPA s? )

3
Cepstral Analysis

• Need to modify sounds into a format meaningful to
  the network
• Mel Frequency Cepstral Coefficients (MFCC) are a
  popular method of feature extraction.
• MFCC take a discrete Fourier transform based on a
  modified scale.

The mel scale
4
Publication of Interest

• Recurrent Neural Networks for Phoneme Recognition
• Takuya Koizumi, Mikio Mori, Shuji Taniguchi, and
  Mitsutoshi Maruya
• Dept. of Information Science, Fukui University,
  Japan
• Applied recurrent neural networks to classify
  phonemes from a Japanese word database

5
Overview of recurrent neural networks

• In contrast with feed-forward networks, recurrent
  neural networks can have cycles.
• This means that the input can be split up among
  multiple time steps.
• In this publication, two types of recurrent
  neural networks were studied.

6
Type 1 RNN
7
Type 2 RNN
8
Benefits of recurrent neural networks

• Feedforward multi-layer neural networks are
  inherently unable to deal with time-varying
  information
• In particular, some consonants are difficult to
  distinguish.

9
Group Classification Scheme

• In addition to having a single network classify
  all phonemes, a two level hierarchy was
  developed
• Classify to which phonetic group a phoneme
  belongs (unvoiced plosives, voiced plosives,
  unvoiced frictaves, voiced frictavesglides,
  nasals, vowels).
• Classify phonemes within a specific phonetic
  group

10
Results

• Overall, recurrent neural networks were superior
  to feed-forward neural networks (MLNN).
• Overall, the group classification scheme was more
  effective than a single RNN.
• In most cases, the Type 1 RNN outperformed the
  Type 2 RNN.
• Training affects weights of all the
  connections in the Type 1 RNN, while it affects
  only part of the connections in the Type 2 RNN

11
Detailed Results
Accuracies () Type 1 RNN Type 2 RNN MLNN
Single Network 84.9 75.1 68.5
Group Classification 91.9 88.1 81.3
Intra-group Recognition (average) 95.2 92.2 89.8
Overall Group Classification Scheme 88.1 -- --
12
Application to Korean Classification Problem

• For unvoiced fricatives, the group to which ? and
  ? belong, the network performed as follows

Type 1 RNN Type 2 RNN MLNN
Accuracy () 87.6 84.0 81.1
13
Questions?