Seminar from Aran Mohammadamini, Post-doc fellow at LIUM

Date: 24/03/2025
Time: 11h00
Place: IC2, Boardroom
Speaker: Aran Mohammadamini
 
 

Using large ASR models for training lightweight models in low-resource and computation-limited languages

Low-resource languages often suffer not only from a lack of language resources but also from limited computational resources. Recent multilingual ASR models reduce the amount of data required to train a practical system for low-resource languages; however, they impose another constraint: the need for substantial computational resources for training and inference. We demonstrate how these large models can be leveraged to generate high-quality speech recognition data for low-resource languages, enabling the training of lightweight models that achieve results close to those of the large models.

Our experiments are conducted on Central Kurdish, which is a low-resource language. The obtained WERs of 8.18 and 20.01 on the Asosoft and Fleurs protocols for the Central Kurdish language using the Seamless large model, and attaining WERs of 7.57 and 22.83 on the same protocols with a 75x smaller model, demonstrates the efficiency of our proposed approach.