Collecting Data for Automatic Speech Recognition Systems in Dialectal Arabic Using Games with a Purpose

摘要

Building Automatic Speech Recognition (ASR) systems for spoken languages usually suffer from the problem of limited available transcriptions. Automatic Speech Recognition (ASR) systems require large speech corpora that contain speech and their corresponding transcriptions for training acoustic models. In this paper, we target the Egyptian dialectal Arabic. As other spoken languages, it is mainly used for spoken rather than writing purposes. Transcriptions are usually collected manually by experts. However, this proved to be a time-consuming and expensive process. In this paper, we introduce Games With a Purpose as a cheap and fast approach to gather transcriptions for Egyptian dialectal Arabic. Furthermore, Arabic orthographic transcriptions lack diacritizations, which leads to ambiguity. On the other hand, transcriptions written in Arabic Chat Alphabet are widely used, and include the pronunciation effects given by diacritics. In this work, we present the game Maiameio (pronouced as makhamekho) that aims at collecting transcriptions in Arabic orthography, as well as in Arabic Chat Alphabet. It also gathers mappings of words from Arabic orthography to Arabic Chat Alphabet.