Arab Voices: Mapping Standard and Dialectal Arabic Speech Technology

TL;DR

Substantial heterogeneity is found both in acoustic conditions and in the strength and consistency of dialectal signals across datasets, underscoring the need for standardized characterization beyond coarse labels in DA ASR.

Abstract

Dialectal Arabic (DA) speech data vary widely in domain coverage, dialect labeling practices, and recording conditions, complicating cross-dataset comparison and model evaluation. To characterize this landscape, we conduct a computational analysis of linguistic ``dialectness'' alongside objective proxies of audio quality on the training splits of widely used DA corpora. We find substantial heterogeneity both in acoustic conditions and in the strength and consistency of dialectal signals across datasets, underscoring the need for standardized characterization beyond coarse labels. To reduce fragmentation and support reproducible evaluation, we introduce Arab Voices, a standardized framework for DA ASR. Arab Voices provides unified access to 31 datasets spanning 14 dialects, with harmonized metadata and evaluation utilities. We further benchmark a range of recent ASR systems, establishing strong baselines for modern DA ASR.

Figures (9)

• Figure 1: Mapping framework for standard and dialectal Arabic speech technology. We unify public and licensed datasets through a standardization pipeline, and then profile their demographic and quality characteristics, validating results through human-in-loop review. The unified data are then used to benchmark speech models and multimodal LLMs; the results of which alongside our dataset profiling, helps informs future data collection and modeling efforts.
• Figure 2: WER distribution across languages for the best-performing ASR models from the three architectures listed in Table \ref{['tab:asr_results_wer']}. Stacked bar charts show the proportion of samples in different WER ranges (from $\leq$10% to $>$100%) across multiple languages. Colors range from green (low error rates) to red (high error rates). "OmniASR LLM 7B" consistently achieves a higher share of low-WER samples and more stable performance across languages, highlighting the advantages of LLM-based architectures over CTC-based approaches for multilingual ASR. For the full analysis on WER and CER performance, details are provided in Figure \ref{['appdx_fig:WER_analysis']} and Figure \ref{['appdx_fig:CER_analysis']} Appendix \ref{['appdx_sec:results']}.
• Figure D.1: We plot the relationship between the four audio quality metrics (see Appendix \ref{['appdx_sec:analysismethods']} for detailed descriptions of each metric).While SI-SDR, PESQ, and STOI (shown as color) largely align, we note that this is not the case with the NMR-MOS model (shown as size).
• Figure D.2: Percentage distribution of age ranges in Datasets with age information of speakers. Speakers binned to 10 year age intervals.
• Figure D.3: Percentage of Male (blue), Female (red) and unknown (black) speakers where datasets report gender.