Open Internet by MindsNet

Convergence Issues in Arabic ASR Model Training

An Arabic Automatic Speech Recognition (ASR) model using a Conformer-small encoder and Transformer decoder struggles to converge during training on a 100-hour dialectal Arabic speech dataset. Despite adjusting hyperparameters and modifying the loss functions, the model fails to properly converge, resulting in high validation Word Error Rate (WER). The issue persists even with changes to learning rate, batch size, and vocabulary size.

Computing & Technology, Computer Science, Machine Learning