Preserving the Iranian Turkic Language: Community-Driven ASR Datasets and Benchmarking for South Azerbaijani
Year: 2026
Type: conference-paper
Abstract: Automatic speech recognition (ASR) is widely used, but its progress in low-resource languages like South Azerbaijani, a Turkic language with Arabic script, remains limited due to scarce data.To address this gap, we introduce the first community-driven ASR datasets for South Azerbaijani.These include: a dataset comprising +255,743 transcribed utterances from multiple sources; a dataset collected by the authors from the South Azerbaijani-speaking community; and the AZB ASR GoldSet, a challenging benchmark dataset for future evaluation.To establish strong baselines, we benchmark eight models, including MMS-1b-all and seven Whisper-family models finetuned on our datasets.Our results demonstrate that using the full dataset improves performance, and language-specific finetuning is crucial for low-resource settings.Error analysis further reveals that Arabic script ambiguity and deletion-heavy decoding are the primary error sources.The code and datasets are also publicly available 1 2 .
Source: Interspeech (USB)
Institutions: Amirkabir University of Technology, King's College London
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225140637