aranemini/central-kurdish-pseudolabel

Central Kurdish → English Pseudo-Labeled Speech Translation Corpus

Dataset Summary

This repository contains a large-scale pseudo-labeled speech translation corpus for Central Kurdish (Sorani Kurdish).

The dataset was automatically generated using a pipeline composed of:

1. Speech segmentation
2. Automatic Speech Recognition (ASR)
3. Machine Translation (MT)

The objective is to provide training data for end-to-end Speech-to-Text Translation (S2TT) in a language with very limited parallel speech translation resources. The corpus was created from publicly available Kurdish speech recordings and automatically translated into English.

Dataset Statistics

Property	Value
Source language	Central Kurdish (ckb)
Target language	English (en)
Hours of speech	~3,200 h
Number of samples	~1.7 million
Label type	Pseudo-labeled

The pseudo-labeling pipeline generated approximately 3,200 hours of Kurdish speech aligned with English translations, making it one of the largest speech translation resources available for Kurdish.

Data Quality

The pseudo-labels were generated using state-of-the-art Kurdish speech and language technologies.

The Automatic Speech Recognition (ASR) system employed in the pipeline achieved:

Evaluation Set	WER (%)
Asosoft Test	8.18
FLEURS Test	20.31

These results represented state-of-the-art performance for Central Kurdish ASR at the time the corpus was created and significantly reduced transcription noise during pseudo-label generation.

Results Obtained Using This Corpus

The pseudo-labeled corpus was subsequently used to train lightweight speech processing models for Central Kurdish.

Automatic Speech Recognition

A lightweight ASR model trained using pseudo-labeled data achieved:

Evaluation Set	WER (%)
Asosoft Test	7.57
FLEURS Test	22.83

Despite being substantially smaller than multilingual foundation models, the resulting model achieved competitive performance while offering significantly lower computational requirements.

Speech Translation

An end-to-end Speech-to-Text Translation model trained using the pseudo-labeled corpus achieved the following performance:

Evaluation Set	BLEU	ChrF++
FLEURS Test	21.97	52.52
Asosoft Test	27.23	56.53
COMMUTE Dev	25.73	51.81
COMMUTE Test	21.09	49.26

These results demonstrate that large-scale pseudo-labeling can effectively compensate for the scarcity of manually annotated speech translation corpora in low-resource languages.

Intended Uses

This dataset can be used for:

• End-to-End Speech Translation (S2TT)
• Semi-supervised Speech Translation
• Speech representation learning
• Low-resource speech processing
• Pretraining multilingual speech models
• Research on pseudo-labeling strategies
• Data augmentation for speech translation systems

Limitations

The labels are automatically generated and may contain:

• ASR transcription errors
• Machine translation errors
• Segmentation inaccuracies

Users are encouraged to validate data quality for their specific use cases and, when possible, combine this corpus with manually annotated resources.

Citation

If you use this dataset, please cite:

@inproceedings{mohammadamini25_interspeech,
  title     = {Scaling pseudo-labeling data for end-to-end low-resource speech translation (the case of Kurdish language)},
  author    = {Mohammad Mohammadamini and Aghilas Sini and Marie Tahon and Antoine Laurent},
  booktitle = {Interspeech 2025},
  year      = {2025},
  pages     = {898--902},
  doi       = {10.21437/Interspeech.2025-887}
}

@inproceedings{mohammadamini2026iwslt,
  title        = {LIUM Submission for IWSLT 2026 Low-Resource Speech Translation Track},
  author       = {Mohammad Mohammadamini and Marie Tahon},
  year         = {2026},
  howpublished = {Proceedings of the International Conference on Spoken Language Translation (IWSLT) 2026},
}

@inproceedings{mohammadamini2025frugal,
  title={Apprentissage de modèles frugaux pour les langues peu dotées à partir de larges modèles d'ASR},
  author={Mohammadamini, Mohammad},
  school={Le Mans Université},
  year={2025}
}