Evaluation of speech unit modelling for HMM-based speech synthesis for Arabic

Abstract

This paper investigates the use of hidden Markov models (HMM) for Modern Standard Arabic speech synthesis. HMM-based speech synthesis systems require a description of each speech unit with a set of contextual features that specifies phonetic, phonological and linguistic aspects. To apply this method to Arabic language, a study of its particularities was conducted to extract suitable contextual features. Two phenomena are highlighted: vowel quantity and gemination. This work focuses on how to model geminated consonants (resp. long vowels), either considering them as fully-fledged phonemes or as the same phonemes as their simple (resp. short) counterparts but with a different duration. Four modelling approaches have been proposed for this purpose. Results of subjective and objective evaluations show that there is no important difference between differentiating modelling units associated to geminated consonants (resp. long vowels) from modelling units associated to simple consonants (resp. short vowels) and merging them as long as gemination and vowel quantity information is included in the set of features.

DNN-Based Speech Synthesis for Arabic: Modelling and Evaluation

“…This problem has been dealt with in [7] for HMM-based Arabic speech synthesis, where four modelling approaches are proposed; differentiating geminated consonants (resp long vowels) from simple consonants (resp short vowels) or merging them:…”

“…This section summarizes the experiments described in [7], which were conducted to compare the four modelling approaches listed above in the framework of HMM-based synthesizer. The speech data used to train the speaker-dependent models with HTS was extracted from the corpus developed in [6].…”

“…Later, STRAIGHT vocoder [11] was used in [12] to generate a higher-quality Arabic speech. [7] focused on phonological particularities of Modern Standard Arabic (MSA) [2]. Two phenomena were highlighted, namely gemination [16] (i.e.…”

“…a long vowel is twice as long as its short counterpart). In [7] subjective and objective evaluations showed that considering the geminated consonants (resp long vowels) as fully-fledged phonemes or as the same phonemes as their simple (resp short) counterparts leads to similar speech quality as long as the information about gemination and vowel quantity are included in the set of contextual features.…”

“…Results showed that DNN outperformed HMM in terms of speech quality and naturalness of produced speech for English language [23,19]. This paper aims at introducing DNN in parametric speech synthesis for Arabic and investigating if DNN benefit from the explicit differentiation of different phoneme classes unlike HMM [7]. The paper is organised as follows.…”