Towards a High-Quality Lemma-Based Text to Speech System for the Arabic Language

Recent numbers put the Arabic language at around 250 million native speakers, making it the fifth spoken language regarding the number of speakers. Therefore, it has gained the interest of researchers in speech technologies in particular speech recognition and speech synthesis. Indeed, many researchers are still investigating in Arabic Text To Speech to deliver an intelligible and close to natural Text To Speech systems. Nevertheless, the most of the available free and semi-free Arabic Text To Speech systems are still away from the natural sounding as human voice does, and the generation of smooth voice is still involved. The primary intention of this work is to increase the quality of the produced speech resulting from the sub-segment based approach proposed in our previous work. To this end, a lemma-based approach for concatenative TTS synthesis is adopted and presented in this paper. In this context, a study of Arabic lemmas frequency was conducted to identify the highly frequent lemmas that often occur in written and spoken Classical and Modern Standard Arabic (MSA). This study reports an analysis of roughly 65 million words fully vocalized obtained from Tashkila corpus, Nemlar, and Al Jazeera. These latter cover modern and classical Arabic languages. As a result, an Arabic lemmatized frequency list was generated. The top 1,000 frequent lemmas were found to provide approximately 79% coverage of the Arabic words. Thus, the former were used as the basic acoustic units of our Text to Speech System. Finally, we demonstrate that this approach affords an improvement in the intelligibility and naturalness of a Text To Speech system with an overall rate 4.5 out of 5.