Arabic Little STT: Arabic Children Speech Recognition Dataset

AI summary

Overview

Research area: Natural Language Processing / Automatic Speech Recognition (ASR) for low-resource and underrepresented speaker groups. Technical level: Intermediate — the paper involves an ASR model evaluation (Whisper variants and WER), but the core message about data scarcity and ethical data governance is accessible to a general audience. Scope: One sentence: The paper introduces Arabic Little STT, a Levantine Arabic children's speech dataset, and reports how Whisper models perform on it relative to adult Arabic benchmarks.

What This Paper Is About

AI speech systems depend on high-quality training data, but Arabic is a low-resource language, and speech from children is especially scarce in existing corpora. This paper addresses that gap by releasing a dataset of Levantine Arabic child speech collected in classrooms and then testing how well a leading ASR model, Whisper, handles it. The goal is to document how far current speech recognition falls short on Arabic-speaking children and to argue for dedicated child speech benchmarks and inclusive training data.

Key Contributions

1. A new child speech dataset: The authors created Arabic Little STT, containing 355 utterances from 288 children aged 6 to 13, recorded in classrooms in Levantine Arabic.
2. A systematic evaluation of Whisper on child speech: They assessed eight Whisper variants on the dataset and compared performance against adult Arabic benchmarks.
3. Evidence of a large child-versus-adult performance gap: The evaluation shows a stark contrast in word error rate between child and adult speech for the same model family.
4. A call for ethical and inclusive data practices: The paper emphasizes that child speech data must be governed by strict ethical and privacy frameworks, and offers the dataset publicly to improve children's demographic representation in ASR data.

Main Findings

• Whisper struggles on Arabic child speech: Even the best-performing variant tested, Large_v3, achieved a 0.66 word error rate (WER) on the child speech data.
• A sharp contrast with adult performance: The same model achieves a sub-0.20 WER on adult Arabic datasets, according to the abstract.
• The gap is not unique to Arabic: The authors note these results align with other research on English speech, suggesting a broader pattern in ASR performance on children.
• Dedicated benchmarks are needed: The results are presented as evidence that child speech benchmarks and inclusive training data are critical for ASR development.
• Ethical safeguards are required: The paper stresses that such data collection and use must be governed by strict ethical and privacy frameworks to protect sensitive child information.

Methodology in Plain English

The researchers built their own dataset to fill a gap they identified: they recorded Levantine Arabic child speech in classroom settings, gathering 355 utterances from 288 children between the ages of 6 and 13. With that dataset in hand, they ran a systematic assessment of Whisper, a state-of-the-art ASR model, across eight of its variants. To judge whether the results were specifically about children or about the model in general, they compared the child-speech performance against the model's performance on adult Arabic benchmarks. The abstract does not describe the recording equipment, annotation process, train/test splits, or the exact decoding settings used — those details are not provided in the abstract.

Why This Matters

• Research impact: The paper documents a severe performance gap on a speaker group that is systematically underrepresented in speech corpora, reinforcing findings from English and pointing to a general shortcoming in ASR development rather than a language-specific one.
• Real-world applications:
  • Educational technology, such as speech-enabled reading or language-learning tools for Arabic-speaking children in classrooms.
  • Voice interfaces and assistants that need to understand younger speakers, who are poorly served by models trained mostly on adults.
  • Accessibility tools, such as speech-to-text support for children with speech or writing difficulties.
  • Speech therapy and language assessment, where accurate recognition of child speech could support screening and progress tracking.
• Industry relevance: Companies building ASR products for Arabic-speaking markets currently have no strong benchmark for child speech; this work signals that adult-benchmark accuracy can substantially overstate real-world performance for younger users, and that ethical handling of children's voice data is a design constraint, not an afterthought.

Future Directions

• Extending the dataset: The abstract presents this collection as an initial step, implying larger and more varied child speech corpora are needed; the abstract does not specify planned scale.
• Adapting or retraining ASR models for child speech: Given Whisper's high error rate on children, follow-up work could explore fine-tuning or child-focused modeling approaches.
• Governance and privacy frameworks: The paper calls for strict ethical and privacy safeguards, raising open questions about consent, storage, and use of children's voice recordings.
• Broadening language and demographic coverage: The work focuses on Levantine Arabic; whether the same gap appears across other Arabic dialects and child populations remains an open question.
• Benchmark standardization: The authors argue for dedicated child speech benchmarks, leaving the design of such standards as future work.

Target Audience

Researchers and practitioners in speech recognition and low-resource language technology; dataset creators and linguists working on Arabic; developers building voice products for children; and policymakers or ethics reviewers concerned with the governance of children's speech data and equitable speech technology.

Authors’ abstract

The performance of Artificial Intelligence (AI) systems fundamentally depends on high-quality training data. However, low-resource languages like Arabic suffer from severe data scarcity. Moreover, the absence of child-specific speech corpora is an essential gap that poses significant challenges. To address this gap, we present our created dataset, Arabic Little STT, a dataset of Levantine Arabic child speech recorded in classrooms, containing 355 utterances from 288 children (ages 6 - 13). We further conduct a systematic assessment of Whisper, a state-of-the-art automatic speech recognition (ASR) model, on this dataset and compare its performance with adult Arabic benchmarks. Our evaluation across eight Whisper variants reveals that even the best-performing model (Large_v3) struggles significantly, achieving a 0.66 word error rate (WER) on child speech, starkly contrasting with its sub 0.20 WER on adult datasets. These results align with other research on English speech. Results highlight the critical need for dedicated child speech benchmarks and inclusive training data in ASR development. Emphasizing that such data must be governed by strict ethical and privacy frameworks to protect sensitive child information. We hope that this study provides an initial step for future work on equitable speech technologies for Arabic-speaking children. We hope that our publicly available dataset enrich the children's demographic representation in ASR datasets.