Modern Arabic speech corpus for Text to Speech synthesis

Abstract

there are hardly any open access large single speaker corpora that could be effectively used to build a Text-to-Speech system, especially for Arabic being a low-resourced language. Thus, the aim of this paper is to present a new open access single speaker corpus. It is by far the largest Arabic speech resource suitable for building text-to-speech systems. The released corpus consists of more than 16-hours audio files aligned with their corresponding phonetic transcription in Buckwalter format, and the orthographic text transcripts representing 81, 000 words fully diactritized. The corpus design was determined by different factors among which is the coverage of the most frequent lemmas, this latter being a common unit in most of the Arabic words. The corpus is freely available for download from http://oujda-nlp-team.net/en/corpora/speech-corpus-1-0/.