EUVIMEDEUVIMED entdecken
Crossref

A DeepSpeech2-Inspired Convolutional Recurrent Framework for Low-Resource Urdu Speech Recognition

Syed Azeem Inam, Syed Nazia Ashraf, Hassan Hashim, Syed Wajiha Naim, Muhammad Ahmed Ansari, Ahmed Raza Khanzada

VFAST Transactions on Software Engineering · 25.04.2026 · 14 (2) · 81-95

DOI: 10.21015/vtse.v14i2.2392 ↗

Kurz gefasst

Fachartikel in der Zeitschrift VFAST Transactions on Software Engineering, online erschienen am 25.04.2026. Verfasst von Syed Azeem Inam und 5 weiteren Autor:innen. Die Arbeit nennt 69 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Low-resource language automatic speech recognition is a difficult task due to small annotated corpora, large speaker and phonetic diversity, and the lack of strong end-to-end metrics. The case of Urdu is especially significant because of the high number of speakers and the inability to provide high-performing open automatic speech recognition systems to date. The study presents an end-to-end Urdu speech-to-text model, built upon a DeepSpeech2-inspired convolutional recurrent neural network, which integrates a spectrogram-based acoustic modeling, bidirectional gated recurrent units, and Connectionist Temporal Classification to learn alignment-free transcription. This model was trained and tested on the Urdu subset of the Mozilla Common Voice corpus with 58,119 training utterances and 6,458 validation utterances and evaluated on a held-out test set. The proposed system has shown to converge consistently during training with a validation Word Error Rate of 21.29% and loss of 5.87 at epoch 478, and a final test Word Error Rate of 17.05, Sentence Error Rate of 34.72, and Word Information Loss of 0.41. The proposed model achieved better performance on the same evaluation setting compared with a reduced recurrent baseline, a transformer-based baseline, and a wav2vec2-style baseline, whose WERs were 23.84%, 19.62%, and 18.31%, respectively. Analysis of ablation also indicated that convolutional feature extraction, as well as deep bidirectional temporal modeling, are essential to performance, and error analysis revealed phonetic confusion, dialectal variation, noise, and high-speed speech as the most prevalent causes of recognition error. These results demonstrate that a well-tuned convolutional recurrent model can provide a competitive solution for Urdu automatic speech recognition under low-resource conditions and offers a reproducible reference point for future studies.

Zitieren

APA

Inam, S. A., Ashraf, S. N., Hashim, H., Naim, S. W., Ansari, M. A., & Khanzada, A. R. (2026). A DeepSpeech2-Inspired Convolutional Recurrent Framework for Low-Resource Urdu Speech Recognition. VFAST Transactions on Software Engineering, 14(2), 81-95. https://doi.org/10.21015/vtse.v14i2.2392

Vancouver

Inam SA, Ashraf SN, Hashim H, Naim SW, Ansari MA, Khanzada AR. A DeepSpeech2-Inspired Convolutional Recurrent Framework for Low-Resource Urdu Speech Recognition. VFAST Transactions on Software Engineering. 2026;14(2):81-95. doi:10.21015/vtse.v14i2.2392

BibTeX
@article{Inam2026,
  author = {Inam, Syed Azeem and Ashraf, Syed Nazia and Hashim, Hassan and Naim, Syed Wajiha and Ansari, Muhammad Ahmed and Khanzada, Ahmed Raza},
  title = {A DeepSpeech2-Inspired Convolutional Recurrent Framework for Low-Resource Urdu Speech Recognition},
  journal = {VFAST Transactions on Software Engineering},
  year = {2026},
  volume = {14},
  number = {2},
  pages = {81--95},
  doi = {10.21015/vtse.v14i2.2392}
}
Kostenlos starten →