A Large Arabic Broadcast News Speech Data Collection

This paper describes the collection and transcription of large amounts of Arabic broadcast news speech data. More than 4000 hours of satellite data have been collected from various Arabic sources. The data was recorded from selected Arabic TV and radio stations in both Modern Standard Arabic and dialectal Arabic. Also, close to 2400 hours of data from a large selection of Arabic sources were transcribed. The orthographic transcriptions used the Quick Rich Transcription (QRTR) method and included segmentation, speaker turns, topics, sentence unit types and minimal noise mark-up. A validation procedure was also established to apply a quality control measure to the transcripts.