Publications/Speech Processing, Recognition & Audio Security
Conference · 2026
Dissecting ASR Failures in Low-Resource South Asian Languages
Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Maryam Mustafa, Agha Ali Raza
Interspeech 2026, Sydney, Australia · CORE A
Summary
Speech recognition now claims to cover more than a hundred languages, but how well does it work for the 350 million speakers of Urdu, Punjabi, Pashto, and Sindhi? We tested five multilingual models on two benchmarks and sorted every mistake into an 11-category error taxonomy. Much of the failure is not about hearing at all: models write in the wrong script, mix in neighboring languages (up to 35% of output characters), and stumble on rare words. Even the best system, SeamlessM4T, reaches only a 16.3% word error rate on Urdu. Simple fixes help: transliteration post-processing recovers up to 25 points of word error rate without any retraining.
Cite this paper
Abdul Hameed Azeemi, Ihsan Ayyub Qazi, Maryam Mustafa, Agha Ali Raza. Dissecting ASR Failures in Low-Resource South Asian Languages. Interspeech 2026, Sydney, Australia.
@inproceedings{azeemi2026dissecting,
author = {Abdul Hameed Azeemi and Ihsan Ayyub Qazi and Maryam Mustafa and Agha Ali Raza},
title = {{Dissecting ASR Failures in Low-Resource South Asian Languages}},
booktitle = {Interspeech 2026, Sydney, Australia},
year = {2026},
url = {https://aghaaliraza.com/publications/2026--Interspeech--Dissecting_ASR_Failures_in_Low-Resource_South_Asian_Languages.pdf}
}
- WER We Stand: Benchmarking Urdu ASR ModelsCOLING 2025, Abu Dhabi, UAE
- Conversations in the Wild: Data Collection, Automatic Generation and EvaluationComputer Speech & Language, vol. 89, 2025
- Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio DatasetFindings of ACL 2024, Bangkok, Thailand
- Speech Emotion Recognition for the Urdu LanguageLanguage Resources and Evaluation (LRE), 2023
- Using Self-Attention DNNs to Discover Phonemic Features for Audio Deep Fake DetectionASRU 2021, Cartagena, Colombia