阿尔巴尼亚语文本转语音外文文献

IFAC PapersOnLine 49 29 2016 259 262 ScienceDirectScienceDirect Available online at 2405 8963 2016 IFAC International Federation of Automatic Control Hosting by Elsevier Ltd All rights reserved Peer review under responsibility of International Federation of Automatic Control 10 1016 j ifacol 2016 11 063 2016 IFAC International Federation of Automatic Control Hosting by Elsevier Ltd All rights reserved 1 INTRODUCTION Text to speech means the conversion of texts into clearly understandable speech This helps the transferring of information from machine to a human being Bickley et al 1998 The speech construction based on written texts which is known as TTS text to speech although has been improved continuously yet it hasn t been solved Various technologies which offer speech synthesis are used such as concatenate synthesis formant synthesis and articulatory synthesis Certainly these three types of technologies have their pros and cons This kind of synthesis requires high quality naturalness and intelligibility So in all systems the aim is to reach the perfection of both characteristics Moreover the technologies are affected by concrete languages where the written text is required to be synthesized In this way taking into consideration the differences in language the speech synthesis for any new language is a peculiarity and a universal application has not been found Especially languages that belong to financially weak countries that have a small impact factor have a great stagnation in this aspect and unfortunately the Albanian language is one of them For solving this issue for the poor country languages we can use TTS for other languages but of course by adopting it The goal of this research is the measurement of opportunities whereas as example was taken the text to speech for Turkish language for reading the texts in Albanian one whereas the adjustment of the text will be done through normalization application made in C programming language The research is based on three main steps The first step is finding the language that has similarities with the primary language which in this case is the Albanian language The second step is letter replacement based on the pronunciation The final step is using creating any application with TTS functions that can read installed speakers and can make the letter replacement in correct order 2 THE CLASSIFICATION OF ALPHABET LETTERS The classification of alphabet letters is specific for each language In our case the Turkish language has 29 letters 22 of them Latin and 7 diacritic The Albanian language has 36 letters where beside Latin letters there are double letters and diacritic ones as well 2 1 Albanian Language As the majority of the world and European languages the Albanian language as well had compiled its alphabet based in the Latin language The Latin alphabet as we know consists of 26 letters such as a b c c d e f g h i j k l m n o p r s t u v x y and z All these letters are used in written English language Italian and also in German and French with some adjustments The combination of graphemes for the formation of special phonemes is not counted as real letters For example the English language is known as a refined language where for interpreting 40 essential sounds are used 200 letter combinations In contrast of these complex ways of writing the Albanian language uses the Latin alphabet but tailored to its phonemic inventory For interpreting all phonemes of spoken Albanian language a total of 36 letters are used Since the Latin alphabet has only 26 letters they aren t 17th IFAC Conference on Technology Culture and International Stability October 26 28 2016 Durr s Albania Copyright 2016 IFAC259 Keywords Text Speech Generation Albanian Turkish Language 17th IFAC Conference on Technology Culture and International Stability October 26 28 2016 Durr s Albania Copyright 2016 IFAC259 University of Prizren Ukshin Hoti Prizren Kosovo Tel 37744107945 e mail blerandkoshi Vienna University of Technology Institute for Handling Devices and Robotics IHRT Favoritenstrae 9 E325 A6 Vienna Austria e mail xhevahirbajrami University of Prizren Ukshin Hoti Prizren Kosovo e mail mentor hamiti uni Abstract Nowadays there are ranges of technologies for converting texts into spoken language speech And their common aim is the artificial compilation of natural language speech Anyway such quality converter still doesn t exist for any language much less for financially weak countries As Text to Speech seems to be a great alternative for blind people who can be informed about digital texts independently in any of their languages the research in this field is more than necessary The research is done by using Turkish language TTS for written texts in the Albanian language The Turkish language is based on Latin letters and the reading of them within sentences is similar to the reading in singularity As the Albanian language has the same characteristics as well the Albanian texts can use TTS of the Turkish language 17th IFAC Conference on Technology Culture and International Stability October 26 28 2016 Durr s Albania Copyright 2016 IFAC259 Blerand Koshi Xhevahir Bajrami Mentor Hamiti 17th IFAC Conference on Technology Culture and International Stability October 26 28 2016 Durr s Albania Copyright 2016 IFAC259 Alternative Creation of Text to Speech Technology for the Albanian Language 17th IFAC Conference on Technology Culture and International Stability October 26 28 2016 Durr s Albania Copyright 2016 IFAC259 260 Blerand Koshi et al IFAC PapersOnLine 49 29 2016 259 262 enough to write 36 letters of the Albanian language that s why 25 letters without are filled with other letters This issue was resolved in the Congress of Manastir in November 1908 where the letters of Albanian language were defined and ordered as follows Instituti i Gjuh sis dhe i Let rsis 1995 a b c d dh e f g gj h i j k l ll m n nj o p q r rr s sh t th u v x xh y z zh Thus if we evaluate the graphic system of Albanian language according to general principles we can freely say that is one of the easiest comparing with others alphabets Moreover the problem of writing was solved very properly 36 phonemes of Albanian are interpreted with 36 special letters which are made by 25 simple letters 9 double letters and two letters containing diacritic signs and In this aspect we can say that the basic graphic principle is respected as the same phoneme is introduced in all possible cases using only one special letter simple or double one Instituti i Gjuh sis dhe i Let rsis 2002 2 2 Turkish Language Turkish alphabet is constructed based on Latin letters it consists of 29 letters where letters as b c d f g h j k l m n p r s t v y z are consonants whereas a e i o u and are vowels Also the Turkish language in some cases uses above letters as a e i o u and vowel to explain the length while reading these letters This alphabet represents modern Turkish pronunciation with a high degree of accuracy and specificity It is the current official alphabet and the latest in a series of distinct alphabets used in different eras The current 29 letter Turkish alphabet was established as a personal initiative of the founder of the Turkish Republic Mustafa Kemal Atat rk 2 3 Main Distinctions Between Two Languages Probably the Albanian and the Turkish languages have many differences but since the paper is about TTS only the letters and their pronunciation are considered Even though both languages are based on the Latin language and are very similar to each other in reading not all letters fulfil this criterion Letters which almost are read the same in the two languages are a b d e f g h i k l m n o p r s t u v z Since the Turkish alphabet doesn t have double letters there will be a difficulty in adapting The main distinction that can be mentioned here is the number of letters where the Albanian language has 7 letters more than the Turkish one Letters like and don t resemble with any of Albanian letters On the other side letters like dh the ll and x don t resemble with any of the Turkish language letters 3 THE NORMALIZATION OF TEXT The normalization of text in general represents the process of transforming the text in any certain form according to previously defined rules with the aim of further processing the text with ease Usually the normalization precedes the process for certain destinations Since such cases may include speech synthesis automated language translation the storage of data based on the respective normalized texts etc As examples of normalization the following actions can be mentioned Normalization based on UNICODE The conversion of all letters within text in small letters and capital ones The omission of punctuation marks within text The replacement of abbreviations with full proper names etc Usually while choosing the actions which are used for the normalization of texts except adjustments we also should take in consider the language in which the text was written For different languages there are also compiled different normalization procedures This is conditioned by the specifics of the respective languages For example there are languages that use the Cyrillic alphabet some languages do not read text letter for letter as it is written some languages use characters with diacritical marks etc So the normalization of texts written in the Albanian language will be made according to use its features in Turkish trying maximally to achieve the best possible results 3 1 The Normalization of Written Text in the Albanian Language The differences between the languages normalization of text written in the Albanian language will be made by adapting several related letters with the Turkish language As some of the letters will be read correctly the rest will remain to be the best solution possible During normalization special attention is paid to priority letters Some character pronunciation such as c in Turkish is close to GJ and XH in Albanian and before replacing them with c the first substitution that was made is c with s Other substitutions are shown in Table 1 where the first priority marked letters or rather the letters that should be normalized before others Special characters such as are translated at the beginning in the Albanian language letters and then are adapted to Turkish Also the numbers 0 9 normalization is made in the same way where the numbers will normalize in a digital basis 100 one double zero Table 1 Replacement of Albanian letters with Turkish ones Letter Replacement Letter Replacement c s y xh c j y gj c zh j x s rr r dh v sh th f ll l q Blerand Koshi et al IFAC PapersOnLine 49 29 2016 259 262 261 4 APPLICATION To solve the problem of normalization the application Normalizuesi 1 0 has been created through programming language C in MS Visual Studio The idea here is to have an application that knows which letters have to be replaced in a particular order according to Table 1 Basically it will find a letter in a text and it will replace it with another Besides this Normalizuesi 1 0 has all functions of a TTS application since it will read all TTS voices installed in the operating system In the code there are 4 different regions 1 letter replacement 2 menu codes 3 TTS methods and 4 adding installed voices to a combo box Visual Studio has a library System Speech Synthesis which is used for TTS methods The whole application doesn t exceed 300 lines of code and you can easily advance the code for other languages by creating new methods in order to use it for further researches Fig 1 Application Normalizuesi 1 0 4 1 Text Reader For reading the texts there is a speaker of Turkish language which has been created by huge generators of Text to Speech voices IVONA TTS For the research the voice of IVONA Feliz 2 was downloaded and to carry out this process the Normalizing 1 0 was used due to the option Choose the reader 5 EVALUATION TESTS AND ANALYSIS RESULTS Evaluation is done through one test which is based on the method MOS mean opinion score Goldstein M 1995 The evaluation based on this method uses the assessment method with 1 5 descriptors 1 bad 2 poor 3 fair 4 good 5 excellent As an addition were given the sentences for testing which were chosen in collaboration with Albanian language lecturers and these sentences were considered as meaningfully specific for the people whose mother tongue is Albanian and for those whose isn t but who can speak it well enough Kastrati et al 2014 Table 2 Sentences used in test 1 T gjith f mij t rrinin t qet 2 T dy filluam t ecnim duke aluar n korridor 3 Po djali sa zinte t recitonte harronte vjersh n 4 Me t zbardhur f mij t u nis n n ekskursion 5 Ajo k ndon bukur 6 Duke mos u mjekuar me koh s mundja iu r ndua 7 P rse ke ardhur deri k tu 8 Era e malit e shp rndante k ng n gjithandej 9 E ndoqe ndeshjen mbr m 10 M doli gjumi nga bubullima The test was done by 15 people of different ages mainly between 22 and 35 Regarding their gender 8 of them were male and 7 female and all of them with at least a bachelor degree Before the test we sent them a video with sentences and voice records and after that through the internet they have listened and sent me the results Results are shown in Table 3 with calculated averages for each sentence and the final outcome Table 3 Evaluation results based on sentences Sentence 5 4 3 2 1 Average 1 8 6 1 4 47 2 1 8 6 3 67 3 8 4 3 3 33 4 14 1 4 93 5 13 2 4 87 6 6 8 1 3 33 7 12 3 4 8 8 9 6 4 6 9 8 6 1 3 67 10 1 11 2 1 3 8 4 15 In addition because we used Turkish reader for the research we asked the attendees to write whether they are familiar with the Turkish language In Table 4 is shown the comparison of those who know Turkish and those who don t Table 4 The results between people who know Turkish and the ones who don t Sentence YES NO 1 4 83 4 22 2 3 83 3 56 3 3 5 3 22 4 5 4 89 5 5 4 78 6 3 17 3 45 7 4 83 4 78 262 Blerand Koshi et al IFAC PapersOnLine 49 29 2016 259 262 8 4 67 4 56 9 3 67 3 67 10 4 3 67 4 25 4 08 MOS perceptual result of the method is done by calculating the average for each sentence From the results of Table 3 we can conclude that the quality is generally at a good level Whereas if we take as a basis the recognition or non recognition of the Turkish language Table 4 shows that it does not matter that much since the results are very similar to each other although slightly better results ha