Breaking Walls: Pioneering Automatic Speech Recognition for Central Kurdish: End-to-End Transformer Paradigm
Authors
arXiv
January 1, 2024
Abstract
Develops end-to-end Transformer-based automatic speech recognition for Central Kurdish using a newly constructed 100-hour speech corpus. The study fine-tunes wav2vec 2.0 XLS-R models and combines them with n-gram language models and a Kurdish tokenizer. The best reported configuration achieved 10.0% WER on validation data and 11.8% WER on the AsoSoft test set.
Keywords
Automatic Speech Recognition
Central Kurdish
Sorani
wav2vec 2.0
XLS-R
Transformer
Related Datasets
AsoSoft Speech Corpus
2021
43.68 hours (full corpus); public subset approximately 30 hours
Audio, Text Transcriptions
Central Kurdish (Sorani) speech corpus developed for automatic speech recognition and related speech-processing research. The full corpus reported in the Jira study contains 43.68 hours of speech from …
Central Kurdish 100-Hour ASR Corpus
2024
Approximately 100 hours
Audio, Text Transcriptions
A Central Kurdish (Sorani) speech corpus constructed for end-to-end automatic speech recognition research. The Breaking Walls study reports approximately 100 hours of speech and uses the corpus to …
Citation
Abdullah, A. A., Veisi, H., & Rashid, T. A. (2024). Breaking Walls: Pioneering Automatic Speech Recognition for Central Kurdish: End-to-End Transformer Paradigm. arXiv:2406.02561.