Malay Speech Corpus

CHAPTER 3

MALAY SPEECH CORPUS

3.1 Introduction

The knowledge related to the structure of the rules and grammar for any language must be understood in depth prior to the development of any Automatic Speech Recognition (ASR) systems. This chapter is intended to discuss the related issues concerning the Malay language and its speech sounds.  The Malay corpus and the test collections used for this study are also presented in the following sections.

3.2Malay Speech Sounds and Language Rules

Malay is an Austronesian language spoken by the Malay people who are native to the Malay Peninsula, southern Thailand, Singapore and parts of Sumatra and also known locally as Bahasa Melayu. It is the official language of Malaysia and is an agglutinative language, meaning that the meaning of the word can be changed by adding the necessary prefixes or suffixes that will be explained through out of this section.

The smallest unit in any language is known as phoneme.  The substitution of this unit for another might make a distinction of meaning (Nong et al. 2001).  Integrating the phonemes produces the syllable and words.  Generally, phoneme classification for Malay language is divided into three major groups that consist of Vowels (V), Consonants (C) and other miscellaneous (Manaf & Hamid 1996).  This structure is relatively same with the English language as shown in Figure 3.1 (Karim 1996).

The vowel class comprises of six vowels that is: /a/, //, /i/, /o/, /u/ and /e/.  The vowel sound is produced when the air exit from the lunges and mouth without ant noise.

The second category, which is consonant class, can be further divided into seven different categories that is the stops or plosive group, affricates, nasals, glides, liquids, fricatives and the semivowel.  The sounds from consonants are produced by air from lungs and consist of noise.  The noise is generated in mouth and nose, for instance, phoneme /p/and /b/.  Figure 3.2 describe the consonant utterances classification for the Malay language.

The last category, miscellaneous category, consists of the diphthong and vowel functions.  Vowel function is a combination of two different vowel (ia, io and iu) and most often used in words absorbed directly from its English equivalent such as radio and audio, and in some original Malay words such as nyiur (coconut), hias (decorate) (Hussain, 1997).

3.2.1Malay morphology

Malay morphology is defined as study of word structures in Malay language (Lutfi Abas, 1971). A morpheme is the term used in the morphology. A morpheme is the smallest meaningful unit in a language. In another words, morpheme is a combination of phonemes into a meaningful unit. A Malay word can be comprised of one or more morphemes. When we talk about Malay morphology, we cannot avoid from discussing the process of word formation in Malay language. It is a language of derivative which allows the addition of affixes to the base/root or primary word to form new words. The language itself is different from the English. In English language, the process involves the changes in the phonemes according to their groups. The processes of word formation in Malay language are in the forms of primary words, derivative words, compound words and reduplicative words.

3.2.1.1Primary word

Primary or root words are either nouns or verbs, which is does not take any affixes or reduplication. A primary word can be comprised of one or more syllables. A syllable consists of a vowel (V) or a vowel with a consonant (C) or a vowel with several consonants. The vowel can be presented at the front or back of the consonants. In Malay language, primary word with one syllable accounts for about 500 only (Nik Safiah Karim et al. 1995). Some of the primary words are taken from other languages such as English and Arabic. The structures of the syllable are shown in Table 3.1. Primary words with two syllables are the majority in the Malay language. The structures of the words are shown in Table 3.2 with example of words that illustrated as in Figure 3.3. Primary words with three and more syllables exist in a few numbers. Most of them are taken from other languages as shown in Table 3.3.

Table 3.1:  Structure of words with one syllable

Syllable Structure

Example of word

CV

Ya (yes)

VC

Am (common)

CVC

Sen (cent)

CCVC

Stor (store)

CVCC

Bank (bank)

CCCV

Skru (screw)

CCCVC

Skrip (script)

Table 3.2:  Structure of words with two syllables

Syllable Structure

Example of word

V + CV

Ibu (mother)

V + VC

Air (water)

V + CVC

Ikan (fish)

VC + CV

Erti (meaning)

VC + CVC

Empat (four)

CV + V

Doa (pray)

CV + VC

Diam (silent)

CV + CV

Guru (teacher)

CV + CVC

Telur (egg)

CVC + CV

Lampu (lamp)

CVC + CVC

Jemput (invite)

Read also  Linguistics Essays - William Labov

E

R

+

T

I

V

C

+

C

V

J

E

M

+

P

U

T

C

V

C

+

C

V

C

C – Consonant

V – Vowel

Table 3.3:  Structure of words with three syllables or more

Syllable Structure

Example of word

CV + V + CV

Siapa (who)

CV + V + CVC

Siasat (investigate)

V + CV + V

Usia (age)

CV + CV + V

Semua (all)

CV + CV + VC

Haluan (direction)

CVC + CV + VC

Berlian (diamond)

V + CV + CV

Utara (north)

VC + CV + CV

Isteri (wife)

CV + CV + CV

Budaya (culture)

CVC + CVC + CV

Sempurna (perfect)

CVC + CV + CVC

Matlamat (aim)

CV + CV + VC + CV

Keluarga (family)

CV + CVC + CV + CV

Peristiwa (event)

CV + CV + V + CVC

Mesyuarat (meeting)

CV + CV + CV + CVC

Munasabah (reasonable)

V + CV + CVC + CV + CV

Universiti (University)

3.2.1.2Derivative word

Derivative words are the words that are formed by adding affixes to the primary words. The affixes can exist at the initial (Prefixes), within (Infixes) or final (Suffixes) of the words. They can also exist at the initial and final of the words at the same time. These kinds of affixes are called confixes. Examples of derivative words are “berjalan” (walking), “mempunyai” (having), “pakaian” (clothes) and so on.

3.2.1.3 Compound word

Compound words are the words that are combined from two individual primary words, which carry certain meanings. There are quite lots of compound words in Malay language. Examples of compound words are “alat tulis” (stationery), “jalan raya” (road), “kapal terbang” (aeroplane), “Profesor Madya” (associate professor), “hak milik” (ownership), “pita suara” (vocal folds) and so on. Some of the Malay idioms are from the compound words such as “kaki ayam” (bare feet), “buah hati” (gift), “berat tangan” (lazy), “terima kasih” (thank you) and so on.

3.2.1.4 Reduplicative word

Reduplicative words, as its name suggests, are the words that are reduplicated from the primary words. There are three forms of reduplication in Malay language: full, partial and rhythmic. Examples of reduplicative words are “mata-mata” (policeman), “sama-sama” (welcomed) and so on.

3.3Malay Speech Corpus Design

Malay speech design basically involves the proper selection of speech target sounds for speech recognition.  The Malay phonemes can be analyzed according to the descriptive analysis and distinctive feature analysis.  Generally, the descriptive analysis is preferred over the distinctive feature analysis because it is easier to be implemented. To develop a baseline system for spoken Malay utterances or word model, we need database for isolated spoken Malay words.  However, very little of the literature and reference material in Malay is available in raw electronic form to support research and development work. These materials are sometimes not suitable for the real life speech recognition system due to their setting environments and most of these materials are recorded the planned or read text.Since no spoken Malay database exists, we develop the Malay corpus based on Hansard documents from Parliament of Malaysia. The hansard documents consists of Dewan Rakyat(DR)Parliamentary debates session for the year 2008.  It contains spontaneous and formally speeches and it is the daily records of the words spoken by 222 elected members of DR. The hansard documents comprises of 51 huge raw video and audio files (.avi form) of daily recorded parliamentary session and 42 text files (.pdf form). Each part of parliamentary session contains six to eight hours spoken speeches that surrounded with medium noise condition or environment (less than 30 dB), speakers interruption (Malay, Chinese and Indian) and different speaking styles (low, medium and high intonation or shouting).  The reason of chosen this kind of data is due to their spontaneous and natural way of speaking in a formal or standard Malay speech during the debates session.

The analysis has been done to the whole recorded session from mid-term until the end 2008 of hansard documents. Out of 42 text documents and 51 video files, only 22 text documents and 22 video files were being selected due to their perfect matched in terms of the contents of video and audio source files. The remaining of the text documents and video files have not been chosen due to the missing of some text documents that could not be downloaded, some video files having corrupted during recording session and some of the recorded video having missed sounds.  This study focused and concerned to the video that have audio sounds since it will be used to develop the Malay corpus and to evaluate the performance of isolated spoken Malay speech recognition system. The quantitative information analysis, about the videos and text documents being selected is given in Table 3.4.

Read also  Thought and Language

Table 3.4: Quantitative information of Hansard documents selected.

No.

Video & Text Documents

No. of Topic

No. of Speakers

Total Words

1.

DR28052008 (MEI)

11

129

40,283

2.

DR29052008 (MEI)

15

114

39,612

3.

DR24062008 (JUNE)

13

154

49,212

4.

DR25062008 (JUNE)

10

118

38,053

5.

DR30062008 (JUNE)

10

175

58,013

6.

DR02072008 (JULY)

14

187

67,906

7.

DR03072008 (JULY)

12

120

48,411

8.

DR07072008 (JULY)

16

210

72,890

9.

DR10072008 (JULY)

13

132

42,350

10.

DR28082008 (AUGUST)

10

123

40,780

11.

DR03112008 (NOVEMBER)

17

232

78,750

12.

DR04112008 (NOVEMBER)

11

136

43,440

13.

DR10112008 (NOVEMBER)

10

105

39,560

14.

DR20112008 (NOVEMBER)

16

109

42,795

15.

DR26112008 (NOVEMBER)

10

186

38,880

16.

DR27112008 (NOVEMBER)

10

147

41,450

17.

DR01122008 (DECEMBER)

7

118

38,430

18.

DR02122008 (DECEMBER)

9

176

56,815

19.

DR03122008 (DECEMBER)

12

152

48,616

20.

DR04122008 (DECEMBER)

11

192

56,780

21.

DR10122008 (DECEMBER)

6

130

38,677

22.

DR11122008 (DECEMBER)

10

143

52,369

TOTAL

The process of documents analysis shows that the majority of the Malay words are comprised of primary word with two syllables and mono (one) syllables. Among the Malay words, the syllables structure of VC, CV and CVC are the most common.  These structures are preferred because they are easy to be pronounced exactly as it’s written and their number is quite substantial in the hansard documents. In order to get a good distribution of consonants and vowels for the dataset from the hansard documents, the most frequently primary (root or base) words spoken by speakers during Parliamentary debates are used. As mentioned previously, most of the root words are the primary words that are either in nouns or verbs without adding any derivations (affixes and suffixes) or reduplication to the root words. Thus, from the text documents analysis, we determined 100 primaries words that mostly spoken by the committee members during the debates that consist of 10 primary words of one syllable, four primary words from three or more syllables structures and 86 primary words that form two syllables structures as depicted in Table 3.5. The details quantitative analysis of each words distribution is represented in Appendix A. Each primary word has maximum number of 50 repetitions that uttered by same or different speakers. Thus, there are a total of 5000 isolated spoken Malay words used for this research. The challenging task is to capturing and segmenting the exact words being uttered accordingly to the audio sounds in the video files. The process of creating isolated spoken Malay corpus is illustrated as in Figure 3.4 and briefly explained in the following sections.

Table 3.2: Selection of 100 isolated spoken Malay words as the speech target sounds.

No.

Words

Structures

No.

Words

Structures

1

ADA

V + CV

51

LAGI

CV + CV

2

AHLI

VC + CV

52

LAIN

CV + VC

3

AKAN

V + CVC

53

LAMA

CV + CV

4

AKTA

VC + CV

54

LANGKAH

CVCC + CVC

5

ARAH

V + CVC

55

LEBIH

CV + CVC

6

ATAS

V + CVC

56

MAKLUM

CVC + CVC

7

ATAU

V + CVV

57

MANA

CV + CV

8

BAGI

CV + CV

58

MASA

CV + CV

9

BAIK

CV + VC

59

MASIH

CV + CVC

10

BAKAL

CV + CVC

60

MESTI

CVC + CV

11

BANK

CVCC

61

MUNGKIN

CVCC + CVC

12

BARU

CV + CV

62

NANTI

CVC + CV

13

BEKAS

CV + CVC

63

OLEH

V + CVC

14

BERI

CV + CV

64

ORANG

V + CVCC

15

BINCANG

CVC + CVCC

65

PADA

CV + CV

16

BOLEH

CV + CVC

66

PIHAK

CV + CVC

17

BUAT

CV + VC

67

PRINSIP

CCVC + CVC

18

BUKAN

CV + CVC

68

PULA

CV + CV

19

DALAM

CV + CVC

69

PUN

CVC

20

DAN

CVC

70

RAMAI

CV + CVV

21

DASAR

CV + CVC

71

RIBU

CV + CV

22

DATANG

CV + CVCC

72

RUJUK

CV + CVC

23

DENGAN

CV + CCVC

73

SAH

CVC

24

DIA

CVV

74

SAMA

CV + CV

25

EKONOMI

V + CV + CV + CV

75

SANGAT

CV + CCVC

26

ESOK

V + CVC

76

SAYA

CV + CV

27

HADIR

CV + CVC

77

SEBAB

CV + CVC

28

HAK

CVC

78

SEBUT

CV + CVC

29

HAL

CVC

79

SEDANG

CV + CVCC

30

HARI

CV + CV

80

SEDIA

CV + CVV

31

HENDAK

CVC + CVC

81

SUDAH

CV + CVC

32

IAITU

VV + V + CV

82

SUSAH

CV + CVC

33

IALAH

VV + CVC

83

TADI

CV + CV

34

INGAT

VC + CVC

84

TAHU

CV + CV

35

INGIN

VC + CVC

85

TAHUN

CV + CVC

36

INI

V + CV

86

TIDAK

CV + CVC

37

ISU

V + CV

87

TANYA

CV + CCV

38

ITU

V + CV

88

TELAH

CV + CVC

39

IZIN

V + CVC

89

TENTANG

CVC + CVCC

40

JADI

CV + CV

90

TERIMA

CV + CV + CV

41

JANGAN

CV + CCVC

91

TIDAK

CV + CVC

42

JAWAB

CV + CVC

92

TIPU

CV + CV

43

JUGA

CV + CV

93

TUAN

CV + VC

44

JUTA

CV + CV

94

TUGAS

CV + CVC

45

KABINET

CV + CV + CVC

95

TULIS

CV + CVC

46

KASIH

CV + CVC

96

UNTUK

VC + CVC

47

KAUM

CV + VC

97

WAKIL

CV + CVC

48

KES

CVC

98

WAKTU

CVC + CV

49

KIRA

CV + CV

99

WANG

CVCC

50

KITA

CV + CV

100

YANG

CVCC

Read also  Irregular verbs

3.3.1Corpus Preparation

The Malay corpus creation from Hansard documentsis designed to collect realistic audio data (from the video files) that best represents the actual noise environment in which the autonomous to the parliamentary debates session. All the videos recorded files have the standard of CD audio sampling frequency rate which is chosen to be 44.1 kHz in stereo channel.  However, the human ear is most sensitive to a frequency spectrum ranging from 500 Hz to 4000 Hz, which roughly corresponds to the speech bandwidth carried along analog telephone lines (Marsh 1999).  Thus, in order to capture the voiced and unvoiced sounds of the recorded signals, the audio waveforms are re-sampled at 16 kHz sampling rate and quantized to 16 bits per sample that is needed to be sufficient for this study. The process first begins with manually segments each video into their respective topics debates session of the day accordingly to the text documents which have been selected as perfect matched in the earlier process.  Each video corresponds to the different topics and there are 10 to 15 topics (usul dan pertanyaan) to be debates by the parliamentary committees at the Dewan Rakyat(DR).

In this study, we only concerned to capture the audio signals, thus the process of extracting the video into the waveform signals is needed. This process involved existing shareware software namely Audio Extracted (version 2.3) that has been used as a tool to extracts and convert each video topics into each audio signals. Thus, there are 253 topics in the form of wave files (.wave) that corresponds to the text documents (as depicted in Table 3.4) to be re-sampled at 16 kHz sampling rate and quantized to 16 bits per sample. Finally, the process of re-sampled the waveform signals is done by using a sound editor, Cool Edit Pro, version 2.0.  According to Nyquist theorem (Salam et al. 2000), the original sound can be better replicated when the sampling rate is at least twice in frequency of the original sound.  By using Cool Edit Pro, all the input signals were manually segmented to obtain isolated spoken Malay corpus that will be used for training and testing purposes.

3.3.2Acquisition of Malay Speech Dataset

The acquisition of Malay speech dataset consists of training dataset and testing dataset.  Commonly, there is no general solution to divide the dataset into training and testing dataset (Zhang et al. 1998). However, generally the number of data needed for training is more than for testing. Therefore, in the preliminary study, it is found that neural network classifier with Multi-layer Perceptron (MLP) needs more training data to have good learning and achieve optimal performance. Hence, guidance in selecting the number of training and testing data is done according to the previous research (Salam et al. 2000; Kim & Kim 2001).

It is not the purpose of this research to develop a full scale speech recognizer with large dataset, but to test new techniques by developing a prototype.  Considering this goal, all the approaches were tested on word recognition with medium dataset. All the experiments reported for word recognition accuracy, used 60% of 5000 (3000 words) isolated spoken words for training and the remaining of 40% are used for testing (2000 words) purposes.

3.4Summary

This chapter presents the creation of isolated spoken Malay database or corpus.  The fundamental of Malay language and their rules is also presented as guidance for creating the corpus.  The process of creating Malay corpus is a challenging task since no spoken or naturally speech corpus in Malay language is available.  The process of capturing and segmenting the audio signals is very time consuming, however the manual process could be as foundation or starting point for other researchers to used this database as a reference pattern in speech recognition fields.  Finally, the Malay database will be used at the next level of speech processing stage before proceed further to recognition phase.

Order Now

Order Now

Type of Paper
Subject
Deadline
Number of Pages
(275 words)