Source-linked AI summary
AraYoungVoices: A Diverse L1/L2 Corpus of Arabic Child and Adolescent Speech
Shammur Absar Chowdhury, Zien Sheikh Ali, Houssam Eddine-Othman Lachemat, Hamdy Mubarak
TL;DR
Arabic ASR evidence remains limited for children, adolescents, and L2 speakers, whose speech varies across age, nativity, and dialect. AraYoungVoices pairs a diverse young-speaker read-speech corpus with ASR benchmarks. L2 speech is more challenging, especially for younger speakers, while joint fine-tuning balances performance across age-nativity groups and ASR outputs tend toward the reading prompt.
Problem
ASR performance remains uneven across age and language backgrounds, while adolescents and young L2 speakers remain challenging and comparatively understudied.
Source: §1 · p. 1Method
Results
L2 speech is more difficult than L1 speech, especially for younger L2 speakers; age-specific fine-tuning favors its matched group, while joint fine-tuning offers stronger balance across age-nativity groups.
Source: §4 · p. 4Takeaways & Limitations
The authors position the corpus and analyses as support for more robust and inclusive Arabic ASR for young speakers.
Source: §5 · pp. 4–5The corpus covers read speech and is not always uniformly balanced across age, regional, and L1/L2 groups.
Source: §5 · p. 5
Abstract
from arXiv · showhide
State-of-the-art ASR systems primarily target native adult speech, leading to substantial performance gaps for children, adolescents, and L2 speakers. We introduce AraYoungVoices, a 151.72-hour Arabic read-speech corpus from 286 speakers aged 7--18, comprising AraKids (7--12) and AraTeens (13--18). The corpus includes 146 native Arabic (L1) and 140 second-language (L2) speakers, with native speakers spanning Egyptian, Gulf, Levantine, and North African dialectal backgrounds and L2 speakers representing diverse linguistic backgrounds across the Americas, Asia, Africa, and Europe. We benchmark four pretrained ASR models under zero-shot and fine-tuned settings using unseen-speaker-$\&$-unseen-prompt (USUP) and unseen-speaker-$\&$-seen-prompt (USSP) evaluations. Results show that L2 speech remains substantially more challenging than L1 speech, with the largest errors observed mainly for younger L2 speakers. Age-specific fine-tuning improves the matched age group, while joint fine-tuning provides a stronger balance across populations. ASR hypotheses are also consistently closer to the standard reading prompt than to the verbatim transcription, particularly for L2 speech, suggesting partial normalization of reading deviations.
1. INTRODUCTION
The introduction motivates broader Arabic ASR coverage for young speakers, whose speech varies with age, nativity, and dialect. It presents AraYoungVoices and benchmarks designed to examine recognition across these populations.
1. INTRODUCTION
- Arabic resources for child speech remain limited, and resources focused on adolescence or L2 speech are comparatively fewer.Arabic Little STT, for example, contains 355 utterances from children aged 6–13.Source: §1 · p. 1
- The paper identifies age, L1/L2 background, and dialectal variation as factors that robust Arabic ASR for young speakers must account for.It notes that L2 pronunciation is influenced by first language, proficiency, and other language exposure.Source: §1 · p. 1
- AraYoungVoices contains 151.72 hours of Arabic read speech from 286 speakers aged 7–18, including 146 L1 and 140 L2 speakers.It comprises AraKids (ages 7–12) and AraTeens (ages 13–18).Source: §1 · p. 1
- L1 speakers represent Egyptian, Gulf, Levantine, and North African dialects, while L2 speakers bring linguistic backgrounds from four continents.The corpus is intended to support analysis of age, L1/L2 background, and dialectal variation.Source: §1 · p. 1
- The authors evaluate ASR in zero-shot and fine-tuned settings using USUP and USSP, which differ in whether test prompts are unseen or seen during training.The analysis examines recognition across age, nativity, and dialect, including age-specific and joint fine-tuning.Source: §1 · p. 1
- The contributions include fine-tuned ASR models for young Arabic speech and accompanying annotation guidelines.The paper also presents age- and nativity-aware benchmarks and analyses of recognition gaps.Source: §1 · pp. 1–2
2. ARAYOUNGVOICES
AraYoungVoices is a 151.72-hour Arabic read-speech corpus spanning 286 speakers aged 7–18, with both L1 and L2 speakers and four L1 dialectal backgrounds. Its speaker-disjoint splits and paired prompt–verbatim references support analysis of age, nativity, dialect, and reading deviations.
2. ARAYOUNGVOICES
- The corpus contains 151.72 hours of speech from 286 speakers aged 7–18, divided into AraKids (7–12) and AraTeens (13–18).It includes 146 L1 speakers from Egyptian, Gulf, Levantine, and North African backgrounds and 140 L2 speakers from diverse linguistic backgrounds.Source: §2 · p. 2
- USUP uses unseen speakers and prompts, while USSP uses unseen speakers reading prompts also present in training.The speaker-disjoint USUP test contains 2,500 utterances; USSP uses the same speaker pool with prompts seen during training.Source: §2 · p. 2
- USUP pairs each original reading prompt with a manually produced verbatim transcript, marking substitutions, insertions, deletions, and other speech events.Dialectal and pronunciation-specific realizations are preserved, enabling evaluation and fine-grained analysis of reading and pronunciation variation.Source: §2 · p. 2
- L2 speakers show more shadda drops than L1 speakers (120 vs. 21) and more long-vowel alif omissions (65 vs. 31).For L1 speakers, final short-vowel deletion often reflects pausal pronunciation and loss of grammatical markers rather than reading errors.Source: §2 · p. 3
- Among vowel changes, 42% of L1 changes are vowel→sukun, while 34% of L2 changes move toward fat’ha, with u→a most common (104, 19.5%).The authors characterize the overall contrast as L1 speakers omitting short vowels and L2 speakers regularizing them.Source: §2 · p. 3
- L2 insertions of alif and ya’ suggest vowel lengthening, especially among speakers from the Americas and Asia; L1-Teen insertions often reflect restored diacritics scored as errors.L1-Kids show fewer such diacritic insertions because their prompts are fully diacritized.Source: §2 · p. 3
3. EXPERIMENT SETUP
The experiments compare four Arabic and multilingual speech models under zero-shot and fine-tuned conditions. Evaluation uses USUP and USSP with WER and CER after shared Arabic text normalization.
3. EXPERIMENT SETUP
- Four models are evaluated: AURA-STT (347M), Cohere Transcribe Arabic (2B), Qwen2.5-Omni-Arabic (7B), and Whisper Large-v3 (1.55B).The set includes dedicated Arabic ASR and general-purpose omni models.Source: §3 · p. 3
- Each model is tested zero-shot and fine-tuned on AraKids only, AraTeens only, or joint training data, with speakers kept disjoint from evaluation.Source: §3 · p. 3
3.1. Evaluation Measure
- USUP and USSP are evaluated using WER and CER after the same Arabic normalization is applied to prompts, verbatim transcripts, and hypotheses.Normalization removes diacritics, tatweel, and punctuation, among other processing steps.Source: §3.1 · p. 3
4. RESULTS AND ANALYSIS
Across models, L2 speech is more difficult than L1 speech, and the largest error burden falls mainly on younger L2 speakers. Age-matched tuning favors its target group, while joint tuning offers a stronger balance; hypotheses are closer to prompts than verbatim transcripts.
4. RESULTS AND ANALYSIS
- On USSP, AraKids-only tuning favors AraKids and AraTeens-only tuning favors AraTeens; on the harder USUP set, joint tuning is often competitive or better.The authors describe joint fine-tuning as offering a stronger balance across the full 7–18 age range.Source: §4 · p. 4
- For every USUP setting, WERP is lower than WERV, with a larger difference for L2 speech, indicating hypotheses are closer to prompts than verbatim transcripts.The authors interpret this as partial normalization of reading deviations.Source: §4 · p. 4
- Manual alignment identifies systematic age-, L2-, and dialect-related pronunciation variation alongside model-specific orthographic and decoding errors.The authors say these model-related errors should be distinguished from genuine pronunciation differences when evaluating young-speaker ASR.Source: §4 · p. 4
- L2 speech is consistently more difficult than L1 speech across models, with most errors coming from younger L2 speakers.In zero-shot evaluation, L2 children also have higher WER than L2 adolescents; fine-tuning particularly reduces this gap.Source: §4 · p. 4
5. CONCLUSIONS
ARAYOUNGVOICES benchmarks show substantial L2 recognition gaps, especially for younger speakers, alongside improvements from age-specific and joint fine-tuning. The conclusion also reports prompt-directed normalization and notes limits in the corpus scope and balance.
5. CONCLUSIONS
- L2 speech shows substantial performance gaps, particularly among younger speakers, while age-specific and joint fine-tuning significantly improve recognition.Source: §5 · p. 4
- ARAYOUNGVOICES is, to the authors’ knowledge, the first corpus jointly covering Arabic child and adolescent speech ages 7–18 across L1 and L2 speakers.Source: §5 · p. 4
- ASR systems tend to normalize some reading deviations toward the intended prompt, according to prompt–verbatim analysis.Source: §5 · p. 5
- The corpus is limited to read speech, is not uniformly balanced across groups, and will not publicly release raw audio or full speaker-level metadata.Source: §5 · p. 5