Fleurs-Badini: Translation and Recording Fleurs Dataset for Badini Variant of Northern Kurdish

2026-06-29 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

The FLEURS-Badini dataset is introduced as a dialect-focused extension to the multilingual FLEURS speech benchmark, specifically addressing the underrepresentation of Northern Kurdish (Badini). This new dataset, presented at IWSLT 2026, comprises 5,224 utterances paired with translated text, collected from 45 speakers. Its creation involved a structured process of translation, recording, and validation. Baseline experiments conducted with the dataset reveal significant challenges for automatic speech recognition (ASR) and speech-to-text translation (S2TT) in Badini. The W2V-BERT CTC model achieved the best ASR performance with a Word Error Rate (WER) of approximately 55% on the test set. Speech-to-text translation also showed limited performance, yielding BLEU scores of 6.13 on the development set and 5.24 on the test set. FLEURS-Badini aims to provide a standardized foundation for evaluating speech systems in this dialect.

Key takeaway

For NLP Engineers and Research Scientists working on low-resource languages, you should consider the significant performance gaps highlighted by the FLEURS-Badini dataset. Your current ASR and S2TT models will likely struggle with Badini Kurdish, as evidenced by a 55% WER for ASR and low BLEU scores for S2TT. Focus your efforts on developing robust techniques for highly challenging dialects, potentially exploring novel architectures or more effective data augmentation strategies to improve these benchmarks.

Key insights

Badini Kurdish speech data is critical for advancing ASR and S2TT in underrepresented dialects.

Principles

• Dialect-specific datasets improve multilingual benchmark coverage.
• Low-resource dialects pose significant ASR/S2TT challenges.

Method

The dataset was constructed via a structured process involving translation, recording from 45 speakers, and subsequent validation of 5,224 utterances.

In practice

• Use FLEURS-Badini for Badini ASR/S2TT benchmarking.
• Evaluate W2V-BERT CTC as an ASR baseline.

Topics

• Badini Kurdish
• Speech Recognition
• Speech-to-Text Translation
• Low-Resource Languages
• FLEURS Dataset
• W2V-BERT CTC

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• I spent 4 days training an AI model to understand my country. It barely works — Low-resource NLP projects face significant challenges in data collection, translation, and accurate labeling, often yielding modest baselines.
• Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India — Real-world ASR performance in India requires unscripted data, orthographically-informed metrics, and regional analysis to overcome current benchmark limitations.
• PASQA: Pitch-Accent-Focused Speech Quality Assessment Model Trained on Synthetic Speech with Accent Errors — PASQA accurately assesses pitch-accent correctness in synthetic speech by leveraging a specialized dataset and architectural enhancements.
• LARI Dataset: A Native Portuguese Question Answering Dataset from Brasileiras em PLN — LARI dataset provides a human-validated Portuguese QA resource, addressing a critical language resource gap.
• Hard to Be Heard: Phoneme-Level ASR Analysis of Phonologically Complex, Low-Resource Endangered Languages — Data scarcity, not phonological complexity, primarily drives ASR errors in low-resource, typologically complex languages.
• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.