Shared Task 2: Cross-Lingual ASR for South Asian Languages
South Asia hosts one of the world’s richest concentrations of linguistic diversity, yet many of its languages remain underrepresented in speech technologies. Differences in resource availability, scripts, phonology, and domains pose significant challenges to building robust Automatic Speech Recognition (ASR) systems that generalize beyond high-resource settings. This shared task aims to advance multilingual and cross-lingual ASR for South Asian languages by encouraging the development of accurate, efficient, and generalizable models.
Participants are asked to train a multilingual ASR system using eight South Asian languages drawn from different language families and regions: Balti, Bengali, Dhivehi, Marathi, Nepali, Saraiki, Tamil, and Torwali. These languages use different scripts and represent a spectrum of high, medium, and low-resource conditions. The training data will be obtained from the Mozilla Common Voice v23 dataset, ensuring a uniform and openly available data source.
The task is structured into two subtracks:
Small Model Subtrack: Participants will submit a model with fewer than 400 million parameters.
Large Model Subtrack: Participants will submit a model with more than 900 million parameters.
Both subtracks aim to explore scalability, efficiency, and performance trade-offs in multilingual ASR.
The testing phase comprises (a) zero-shot and (b) limited resource ASR adaptation of the developed multilingual ASR systems for an unseen South Asian language not included in the training set.
Evaluation Metrics: Systems will be evaluated using Word Error Rate (WER) and Character Error Rate (CER) on a per-language basis, with macro-averaged CER as the primary metric.
By combining multilingual ASR, model efficiency, and cross-lingual generalization, this shared task aims to foster inclusive speech technologies and advance ASR research for underrepresented South Asian languages.
All deadlines are 11:59 PM UTC-12:00 (“anywhere on Earth”).
First Call for Participation: 15 December 2025
Training Dataset Available: 15 December 2025
Registration Opens & Submission Details released: 5 January 2026
Registration Deadline: 24 January 2026
Test Dataset Release: 25 January 2026
System Output Submission Deadline: 9 February 2026
Official Results Announced: 11 February 2026
Paper Submission Deadline: 20 February 2026
Notification of Acceptance: 20 March 2026
FAQs coming soon. This section will be updated regularly.
Tafseer Ahmed Khan, Mohammad Ali Jinnah University, Karachi.