Evaluating Forced Alignment in an Under-Documented Shakargarh Punjabi Speech Corpus

Authors

  • Adnan Arif PhD Scholar, School of English Language, Technology and Literature, Faculty of Languages and Literature, Minhaj University Lahore, Pakistan Author

Keywords:

forced alignment; Shakargarh Punjabi; speech corpus; language documentation; low-resource languages

Abstract

This study evaluated the applicability and usefulness of forced alignment for developing a searchable speech corpus of under-documented Shakargarh Punjabi, a local Punjabi variety spoken in Narowal District, Punjab, Pakistan. Positioned at the intersection of corpus linguistics, speech technology, and language documentation, the study addressed the problem that low-resource speech recordings often remain insufficiently transcribed, time-aligned, searchable, and reusable for linguistic analysis. The design adopted was a mixed method that consisted of quantitative descriptive evaluation and qualitative error analysis. Primary spoken data were generated from 24 recordings of 12 native speakers using semi structured interviews, natural conversations and short personal accounts that produced 120 rated utterances (19.79 minutes, 1,565 words). The recordings were subsequently transcribed in Shahmukhi Punjabi; with a consistent Roman transliteration based on Montreal Forced Aligner, annotated in ELAN and manually verified in Praat. Successful utterances were those in which the ABDev remained under 50ms and the segment mismatch was not present. 74 (61.7%) utterances were successfully aligned initially and 96 utterances (80.0%) were successfully aligned after project specific pronunciation-lexicon refinement. The number of OOV cases dropped from 62 to 18 instances while the OOV rate dropped from 3.96 to 1.15 OOV/100 words. The remaining problems primarily were those items which were not present in the standard language, connected speech, and vowel reduction. The findings show that semi-automatic alignment, supported by local lexicon refinement and manual validation, can materially improve the searchability and documentary value of Shakargarh Punjabi speech data.

Author Biography

  • Adnan Arif, PhD Scholar, School of English Language, Technology and Literature, Faculty of Languages and Literature, Minhaj University Lahore, Pakistan

    Adnan Arif is a PhD Research Scholar in English Linguistics at the School of English Language, Technology and Literature, Faculty of Languages and Literature, Minhaj University Lahore. He holds an MPhil in Applied Linguistics from the University of Management and Technology. His research interests include Punjabi/Lahnda dialectology, corpus-driven morphosyntax, sociolinguistics, language variation, language documentation, orthography, natural language processing, and cross linguistic research. His current doctoral work focuses on linguistic variation, documentation, and corpus-driven analysis of Punjabi varieties.

Downloads

Published

2026-10-08 — Updated on 2026-10-29