Phonetic Inventory for an Arabic Speech Corpus

Phonetic Inventory for an Arabic Speech Corpus Nawar Halabi, Mike Wald School of Electronics and Computer Science University of Southampton E-mail: nh2f13@ecs.soton.ac.uk, mw@ecs.soton.ac.uk Abstract Corpus design for speech synthesis is a well-researched topic in languages such as English compared to Modern Standard Arabic, and there is a tendency to focus on methods to automatically generate the orthographic transcript to be recorded (usually greedy methods). In this work, a study of Modern Standard Arabic (MSA) phonetics and phonology is conducted in order to create criteria for a greedy meth-od to create a speech corpus transcript for recording. The size of the dataset is reduced a number of times using these optimisation methods with different parameters to yield a much smaller dataset with identical phonetic coverage than before the reduction, and this output transcript is chosen for recording. This is part of a larger work to create a completely annotated and segmented speech corpus for MSA. Keywords: Phonology, Corpus Design, Corpus Evaluation 1. Introduction Speech corpora for speech synthesisers (namely Unit Selection synthesisers) in different languages have increased in number and vary in size. The size of these corpora is rarely justified in the literature as there is no consensus on what is the minimum length for a speech corpus required to build a synthesiser with a natural voice (Bonafonte et al., 2008; B. Bozkurt, Dutoit, Prudon, D’Alessandro, & Pagel, 2002; Clark, Richmond, & King, 2007; Kominek & Black, 2003; Tao, Liu, Zhang, & Jia, 2008). Bonafonte et al. 2008 produced 10 hours each of male and female Catalan speech targeted at speech synthesis. They claim, in their review, that they have found speech corpora ranging from 3 to 12 hours of speech. Oliveira et al., 2008 recorded 13 hours of speech from each of their 4 subjects, also targeting speech synthesis. They claim that unit selection usually requires 3 to 10 hours of speech, without taking a specific language into consideration. Tao et al., 2008 used 7 hours of recording for the Blizzard Challenge 2008 to build a synthesiser for Mandarin. Some speech corpora are shorter like the “awb” voice produced by Kominek and Black (Kominek & Black, 2003) using the ARCTIC database which consists of 1.4 hours of speech. There is even an attempt to build a speech synthesiser using only an hour of speech for Portuguese (Parlikar & Black, 2012). Many more examples of corpora exist with different lengths which do not cater for redundancy. But what matters more than the speech recording length is the transcript that has been chosen to produce it. The transcript needs to have good phonetic and prosodic coverage of the target language such as the examples of Modern Standard Arabic (MSA) used in this work. 2. Phonetic and Prosodic Coverage Phonetic coverage is the ability of the speech corpus to be used to generate as natural synthetic speech that does not occur in the corpus. The main speech synthesis methods that determine this coverage in this work are concatenative, unit selection synthesis methods …

Content truncated to 500 words