ArEnLD: A balanced multilingual speech corpus for Arabic and English dialect identification.
Description
Dataset Description Overview ArEnLD (Arabic–English Language Dialect dataset) is a balanced, multilingual speech corpus designed for dialect identification, dialect-aware speech recognition, speaker profiling, and cross-lingual transfer learning. The dataset comprises exactly 12,000 one-second speech clips (3.33 hours total), perfectly balanced across six distinct dialect classes (2,000 clips per class). Audio Specifications and File Formats The dataset is provided in two parallel versions to accommodate different research needs: 1. Standard ASR Version (16 kHz): 16,000 Hz sampling rate, 1 channel (mono), 16-bit PCM WAV. Optimized and ready for standard deep learning speech models. 2. High-Fidelity Version (44.1 kHz): 44,100 Hz sampling rate, 2 channels (stereo), 16-bit PCM WAV. Preserves the original broadcast extraction quality. Directory Structure and Naming Convention The dataset is organized into two main directories: ArLD (Arabic Language Dialects) and EnLD (English Language Dialects). Ground-truth labels are inherently encoded in the file name prefixes. Each file follows the format: [Prefix]_[Index].wav (e.g., AG_1.wav to AG_2000.wav). Arabic Dialects (ArLD): • Gulf Arabic: Prefix AG_ (Covers 6 Gulf countries) • Levantine Arabic: Prefix AL_ (Covers 4 Levantine countries) • Maghrebi Arabic: Prefix AM_ (Covers 4 Maghrebi countries) English Dialects (EnLD): • American English: Prefix EAm_ • Australian English: Prefix EAu_ • British English: Prefix EBr_ (Covers Southern/Northern England, Scotland, and Wales) Data Curation and Demographics The raw audio was sourced from 45 publicly available television, web series, and broadcast programs on YouTube (approx. 11 hours for Arabic and 15 hours for English). Speech regions were isolated using Praat to extract clean 20–30 second clips, which were then segmented into non-overlapping 1-second clips using Python. • Speakers: Each class features approximately 25–30 unique speakers. • Demographics: The speaker pool is approximately 40% male and 60% female, spanning children, adults, and elderly age groups. • Quality Control: Speakers' native dialect backgrounds were verified against publicly documented biographical information to ensure linguistic authenticity and minimize code-switching.