Arabic Speech Database

1. Product Overview

The Arabic speech database comprises 899 hours of audio, containing 448,611 sentences. It has a sampling rate of 16,000 Hz and a bit depth of 16 bits. A total of 1,348 speakers were invited to record using their mobile phones in quiet environments. The speakers are primarily from Algeria, Morocco, Saudi Arabia, and Egypt. After transcription and annotation by professional linguists, and following rigorous quality checks, the text accuracy exceeds 95%.

2. Speech Data Transcription

Transcribers render the spoken content based on the audio they hear, striving to ensure that the written text accurately reflects the original pronunciation. General guidelines include: 1) The transcribed text must exactly match the spoken audio—no extra words, missing words, or errors. For example, “اثني ثالثة” should be transcribed as “123,” not “one two three.” 2) Numerals must be transcribed in English Latin form, such as “واحد3.” 3) English text should use lowercase for individual words, while proper nouns and acronyms must be capitalized with spaces, e.g., “W W W dot Google dot com,” “C E O.” 4) All annotations must fully correspond to the actual pronunciation without omissions.

3. Directory Structure and Naming Conventions

Directory Structure

The data directory tree is organized as follows:

└─DOC (textual documentation)
││├─wav_list.txt (audio file list)
││├─content.txt (transcription list)
└─SPEECHDATA (audio folder)
││├─F0004 (speaker ID)
││ ││ F0004W0001.wav (audio file)

Naming Conventions

CORPUS/USAGE/SPEAKER_ID/SPEECH_IDe.g./SPEECHDATA/F0004/F0004W0001.wav

Directory Name · Content · Notes

CORPUS · · Speech database identifier

USAGE · SPEECHDATA · Audio folder

SPEAKER_ID · F0004 · Speaker ID

SPEECH_ID · F0004W0001.wav · Audio file

Data Sample Download

Click to download a sample dataset