The Arabic Speech Database: PADAS

Mohamed Khalil Krichi & Cherif Adnan
The Arabic Speech Database:

PADAS
Mohamed Khalil Krichi
[email protected]
Faculty of Sciences of Tunis/ Laboratory of

Signal Processing/ Physics Department
University of Tunis-Manar
TUNIS, 1060, TUNISIA
Cherif Adnan
[email protected]
Faculty of Sciences of Tunis/ Laboratory of

Signal Processing/ Physics Department
University of Tunis-Manar
TUNIS, 1060, TUNISIA
Abstract
This work describes a construction of PADAS Phonetics Arabic Database Automatically segmented
based on a data-driven Markov process. The use of a segmentation database is necessary in
speech synthesis and recognizing speech. Manual segmentation is precise but inconsistent,
since it is often produced by more than one label and require time and money. The MAUS
segmentation and labeling exist for German speech and other languages but not in Arabic. It is
necessary to modify MAUS for establish a segmental database for Arab. The speech corpus
contains a total of 600 sentences recorded by 3 (2 male and 1 female) Arabic native speakers
from Tunisia, 200 sentences for each.
Keywords: HTK, MAUS, Phonetic Database, Automatic Segmentation.
1. INTRODUCTION
Many researches such as automatic speech recognition or speech synthesis are now based on
database e.g. English [1, 2, 3 and 4]. For obtaining a good result, the database must be
balanced, segmented and reduce the noise (noise in step of record)In order to produce a robust
speaker-independent continuous Arabic, a set of speech recordings that are rich and balanced is
required. The rich characteristic is in the sense that it must contain all the phonemes of Arabic
language. It must be balanced in preserving the phonetics distribution of Arabic language too.
This set of speech recordings must be based on a proper written set of sentences and phrases
created by experts. Therefore, it is crucial to create a high quality written (text) set of the
sentences and phrases before recording them. Any work based on the learning step requires a
database to learn the system and then evaluate it. They are a several international databases in
field of speech such as TIMIT which was developed by DARPA Committee for American English.
And we also find other databases in different known languages, such as French and German,
and unknown, as Vietnamese and Turkish.
For Arabic, we have not found a standard database, but we still found a few references. KACST
[5] database developed by the Institute of King Abdul -Aziz in Saudi Arabia.
1.1 KACST
Indeed KACST created a database for Arabic language sounds in 1997. This database was to
created the least number of phonetically rich Arabic words. As a result, a list of 663 phonetically
rich words containing all Arabic phonemes.
Signal Processing: An International Journal (SPIJ), Volume (8) : Issue (2) : 2014
10
The purpose is used for Arabic ASR and text-to-speech synthesis applications.
KACST produced a technical report of the project Database of Arabic Sounds: Sentences in
2003. The sentences of Arabic Database have been written using the said 663 phonetically rich
words. The database consists of 367 sentences; 2 to 9 words per sentence.
The purpose is to produce Arabic phrases and sentences that are phonetically rich and balanced
based on the previously created list of 663 phonetically rich words [6].
1.2 ALGASD
ALGERIAN ARABIC SPEECH DATABASE (ALGASD) [7] developed for the treatment of Algeria
Arabic speech taking into account the different accents from different regions of the country.
Unavailability and lack of resources for a database audio prompted us to build our own database
to make the recognition of numbers and operations of a standard calculator in Arabic for a single
user. We made 27 recordings of 28 vocabulary words.
Database is the most important tool for multiple domains as speech synthesis or speech
recognition. to provide database a interesting and contains all the acoustic units must have all
the possible linguistic combinations .The quality of the final result of the synthesis is directly
dependent on the quality of recordings made during the development of the acoustic units
therefore a filtering step dictionary is mandatory.
The implementation stages can be summarized as follows:
a)
b)
c)
d)
The choice of dictionary (set of sentences contains several examples of phonemes.).

Sound recording expressions.
Noise reduction.
Segmentation.
2. ARABIC LANGUAGE
Statistics show that it is the first language (mother-tongue) of 206 million native speakers ranked
as fourth after Mandarin, Spanish and English [8].The Arabic language is a derivational and
inflectional language. The original Arabic is the language spoken by the Arabs. In addition, it is
the sacred language of the Koran and Islam. Because the spread of Islam and the spread of the
Qur'an, the language became a liturgical language. It is spoken in 22 countries, while the number
of speakers is more than 280 million [9].
2.1 Alphabet
2.1.1 Consonants
The Arabic alphabet consists of twenty eight consonants (see Table 1) basic, but there are
authors who treat the letter (alif) as the twenty ninth consonant. The (alif) behaves as a long
vowel never found as consonant of the root.
Vowels are not as consonants, they are rarely noted. They are written only to clarify ambiguities
in the editions of the Koran or in the academic literature. Indeed, vowels play an important role in
the Arabic words, not only because they remove the ambiguity, but also because they give the
grammatical function of a word regardless of its position in the sentence. In other words, vowels
have a dual function: one morphological or semantic and the other is syntactic. Arabic has two
sets of vowels, the short one and the other long.
2.1.2
Short Vowels
The short vowels
( , , )
are added above or below consonants. When the consonant has no
vowel, it will mark an absence of vowel represented in Arabic by a silent vowel (
).
11
2.1.3 Long Vowels

Long vowels are long letters, they are formed by a brief vowels and one of the following letters (
,,)
2.2 The Diacritics
Short vowels are represented by symbols called diacritics (see Figure 5). Three in number, these
symbols are transcribed as follows:
The Fetha [a] is symbolized by a small line on the consonant ( / ma / )

Damma the [ u ] is symbolized by a hook above the consonant ( / mu / )
The kasra [i] is symbolized by a small line below the consonant (/ mi / )
A small round o symbolizing Sukun is displayed on a consonant when it is not linked to
any vowel.
2.2.1 The Tanwin

The sign of tanwin is added to the end of words undetermined. It is related to exclusion with
Article determination placed at the beginning of a word. Symbols tanwin are three in number and
are formed by splitting diacritics above, which results in the addition of the phoneme / n /
phonetically:
[an]:
[un] :
[in]:
(
/ AIan / )
( / AIun /)
(/ AIin / )
2.2.2 The Chadda

The sign of the chadda can be placed over all the consonants non initial position. The consonant
which is then analyzed receives a sequence of two consonants identical:
FIGURE 1: Example of a sentence / jalAIabuuna limuddati saAItin / ("They play for an hour").
12
Graphem
es
symbol
symbol
Graphem
es
symbol
Graphem
es
t
T
Z
X
x
d
D
symbol
Graphem
es
s
S
s
d
t
D
?
q
k
l
m
n
h
w
a:
i
i:
u
u:
TABLE 1: Arabic consonant and vowels and their SAMPA code.
3. BALANCED SELECTION OF ARABIC WORDS

The syllabic structures in Arabic are limited in number and easily detectable. Every syllable in
Arabic begins with a consonant followed by a vowel which is called the nucleus of the syllable.
Short vowels are denoted by (V) and long vowels are denoted by (VV). It is obvious that the
vowel is placed in the second place of the syllable. These features make the process of
syllabification easier. Arabic syllables can be classified either according to the length of the
syllable or according to the end of the syllable. Short syllable occur only in CV form, because it is
ending with a vowel so it is open. Medium syllable can be in the form of open CVV, or closed
CVC. Long syllable has two closed forms CVVC, and CVCC. Arabic words are composed at least
by one syllable; most contain two or more syllables. The longest word is combined of five
syllables. Table II illustrates Arabic syllables. Some of the Arabic words are spelled together
forming new long words with 6 syllables like () , or 7 syllables like () . There exist a few
Arabic data suitable for HMM-based synthesis, which should ideally include a large number of
Arabic databases from a single speaker and corresponding phonetic transcriptions.
Syllable
li
cvv
English
meaning
to
fii
in
cvc
qul
say
cvcc
bahr
sea
cvvc
maAl
money
cvvcc
zaArr
visit
cv
Arabic example
TABLE 2: Arabic Syllables Types.
3.1 Corpus Description

Creating phonetically rich and balanced text corpus requires selecting a set of phonetically rich
words, which are combined together to produce sentences and phrases. These sentences and
phrases are verified and checked for balanced phonetic distribution. Some of these sentences
and phrases might be deleted and/or replaced by others in order to achieve an adequate phonetic
distribution [10].The corpus, which we used to build our database, is composed of 200 sentences,
with an average of 5 words per sentence. These sentences contain 1000 words, 2600 syllables,
7445 phonemes including 2302 short vowels and long vowels. These sentences were read at an
13
average speed (from 10 to 12 phonemes/second) by Tunisian speakers, two male and a female.
They were sampled at 16 KHz with 16 bits per sample.
3.2 Corpus Analysis
We have carried out a statistical study of our corpus. Table 3 shows the results of this study. We
can note the following results:
The short vowel [a] and the long vowel [a:] appear with a frequency of 17%, followed by
vowels [i] and [i:] with an occurrence frequency of 14.3%. The vowels [u] and [u:]
represent 7%.
The occurrence of the vowel (short and long) is about 37%.
The most frequent Arabic consonants are: [?] (15%), [n] (6.66%), [l] (6.63%), [m]
(5.59%), etc.
Consonant and
vowels
?
b
t
T
x
/X
G
d
D
r
z
s
S
s
d
t
D
?
g
f
q
k
l
m
n
h
w
j
a
a:
i
i:
u
u:
total
Phoneme
Repetitions
523
102
92
70
19
20
35
39
40
102
35
48
73
18
24
19
24
61
23
61
61
80
260
219
261
123
51
80
400
254
400
28
252
23
3920
%
13,34%
2,60%
2,35%
1,79%
0,48%
0,51%
0,89%
0,99%
1,02%
2,60%
0,89%
1,22%
1,86%
0,46%
0,61%
0,48%
0,61%
1,56%
0,59%
1,56%
1,56%
2,04%
6,63%
5,59%
6,66%
3,14%
1,30%
2,04%
10,20%
6,48%
10,20%
0,71%
6,43%
0,59%
100%
TABLE 3: Occurrence Frequency (%) of Arabic Consonants and Vowels.
14
3.3 Noise Reduction

Degradation of signals by noise is an omnipresent problem [11]. In almost all fields of signal
processing the removal of noise is a key problem. The wavelet transform is striking for its great
variety of different types and modifications. A whole host of different scaling and wavelet
functions (or scaling and wavelet coefficients) provide plenty of possible adjustments and
regulating variables [12]. The audio recordings were noisy with a continuous background noise.
Our goal is to reduce this undesirable component. Figures 2 and 3 shows the time signal before
and after filtering for a particular audio file. We note in particular that the zone of silence
highlighted is closer to zero in the filtered signal in the original signal release.
FIGURE 2: Example for Original Speech of Database File.
FIGURE 3: Example for the Same Speech Denoising.
3.4 Automatic Segmentation

Nowadays, the Practical applications of automatic S&L are
implemented as a statistical search
for a S&L k in a space of all possible S&Ls, which can be formulated as [13,14]:
^
k = arg max k P (k | O) = argmax k
P ( k ) p (O | k )
P (O )
(1)
Where, O is the acoustic observation on the corresponding speech signal. The MAUS system
15
models
P (k )
for each recording O. Each path from the start node to the end node represents a
k
P(k ) p(O | k )
possible
and accumulates to the probability
which is determined by HMMs for
each phonemic segment and a simple Viterbi search through the graph yields the
P(k ) p(O | k )
maximal
.
The Munich Automatic Segmentation (MAUS) system developed by Department of Phonetics,
University of Munich, For more details about the MAUS method refer to [15], [16] and [17].
The purpose is analyzing a spoken utterance. Indeed, input a speech wave and some
orthographic form of the spoken text. The text is parsed into a chain of single words (punctuation
marks are stripped) and passed to a texttophoneme algorithm, which is either rulebased or a
combination of lexicon lookup and fallback to the rulebased system.
FIGURE 4: Processing in MAUS.
3.5 Corpus Segmentation and Labeling

Our continuous speech corpus was segmented and labeled with automatic procedure MAUS.
This software requires as input a file in format "wav" of the sentence to be segmented, and a text
file containing the phonetic transcription of the same sentence. A phonetic file (.par). This file
consists of the list of sentence phonemes with their prosodic characteristics.
16
FIGURE 5: Example MAUS segmentation and labeling taken from the Arabic corpus with SAMPA code.
4. RESULT AND EVALUATION

4.1 Result
A database is a collection of accumulated documents. our database defines as follow:
The files (.wav), four files of transcription (txt, word, phn, textGrid) exist for each sentence of the
corpus, which contains respectively:
The text of the marked sentence (.txt) ;

The associated time aligned word transcription (.word) ;
The associated time aligned phonetic transcription (.phn) ;
Temporal description of each phoneme; start time and end time (.textGrid).
FIGURE 6: Temporal description of each phoneme; start time and end time.
17
4.2 Evaluation
In total, 600 sentences were segmented, 400 sentences for the two speakers (male, 200
sentences for every one), 200 sentences for the third speaker (female). For each segmented 200
sentences, we randomly selected 10 sentences for segmented manually. To do this, we need 6
students in our research laboratory, two for each 10 sentences. The results are summarized in
the following table:
speaker
First male speaker
second male speaker
female speaker
Manual segmentation
99%
99%
99%
Automatic segmentation
94%
94.4%
94.1%
TABLE 4: Evaluation Result.
5. CONCLUSION
This paper reports our work towards developing the PADAS Phonetic Arabic Database
Automatically Segmented based on rich phonetic and balanced speech corpus, which is
automatic segmented with the MAUS system. This work includes creating the rich phonetic and
balanced speech corpus; building an Arabic phonetic dictionary, reducing noise by wavelet
method and an evaluation of the automatic segmentation. The current release of our database
contains 1 female and 2 male voices. The purpose of this work is to build a database to be used
in all area of Speech processing. This variety is useful when used in speech synthesis or speech
recognition.
6. REFERENCES
[1] A. Black and K. Tokuda, The Blizzard Challenge Evaluating Corpus-Based Speech
Synthesis on Common Datasets, in Proceeding of Interspeech, Portugal, pp. 77-80, 2005.
[2] S. DArcy and M. Russell, Experiments with the ABI (Accents of the British Isles) Speech
Corpus, in Proceedings of Interspeech 08, Australia, pp. 293-296, 2008.
[3] J. Garofolo, L. Lamel , W. Fisher, J. Fiscus, D. Pallett, N. Dahlgren, and V. Zue, TIMIT
Acoustic-Phonetic Continuous Speech Corpus, Technical Document, Trustees of the
University of Pennsylvania, Philadelphia, 1993.
[4] K. Tokuda, H. Zen, and A.W. Black, An HMM-based speech synthesis system applied to
English, in IEEE Speech Synthesis Workshop, 2002.
[5] M. Alghamdi, A. Alhamid, and M. Aldasuqi, Database of Arabic Sounds: Sentences,
Technical Report, Saudi Arabia, 2003.
[6] M.A. Mansour Kacst arabic phonetics database. Riyadh, Kingdom of Saudi Arabia. 2004.
[7] G.Droua-Hamdani Algerian Arabic speech database (algasd). December 2010.
[8] R. Gordon, Ethnologue: Languages of the World, Texas: Dallas, SIL International, 2005.
[9] A. Omar Dirasat AlSwat AlLugawi.Cairo: Alam Al Kutub 1985.
[10] L. Pineda, G.mez M., D. Vaufreydaz and J. Serignat Experiments on the Construction of a
Phonetically Balanced Corpus from the Web, in Proceedings of 5th International Conference
on Computational Linguistics and Intelligent Text Processing, Lecture Notes in Computer
Science, Korea, pp. 416-419, 2004.
18
[11] L. Hadjileontiadis and S. Panas. Separation of discontinuous adventitious sounds from

vesicular sounds using a wavelet based filter, IEEE Trans. Biomed. Eng., vol. 44, n 7, pp.
876-886, 1997.
[12] S. Mallat. A wavelet tour of signal processing. Academic Press, 1999.
[13] F. Schiel, and J. Harrington: Phonemic Segmentation and Labelling using the MAUS
Technique. Workshop 'New Tools and Methods for Very-Large-Scale Phonetics Research',
University of Pennsylvania, January 28-31, 2011.
[14] F. Schiel, MAUS Goes Iterative. Proc. of the IV. International Conference on Language
Resources and Evaluation, Lisbon, Portugal, pp. 1015-1018. 2004.
[15] J.L. Fleiss Measuring nominal scale agreement among many raters. Psychological Bulletin,
Vol. 76, No. 5 pp. 378-382. 1971.
[16] S. Burger, K. Weilhammer, F. Schiel, H. G. Tillmann, Verbmobil Data Collection and
Annotation. Foundations of Speech-to-Speech Translation (Ed.Wahlster W), Springer,
Berlin, Heidelberg. 2000.
[17] F. Schiel, C. Heinrich, and S. Barfuer Alcohol Language Corpus. Language
Resources,2011.
19

The Arabic Speech Database: PADAS

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

The Arabic Speech Database: PADAS

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Arabic Speech Database: PADAS

Uploaded by

Copyright:

Available Formats

Mohamed Khalil Krichi & Cherif Adnan