In our empirical study, we explore how informative accent folds are for zero-shot ASR performance, using a cross-lingual model trained on various languages.
We fine-tune the XLSR model with specific hyperparameters, including dropout rates and learning rates, aiming to optimize the model for each target accent.
By leveraging raw speech waveforms and a singular cross-lingual model, we aim to improve accent recognition in Automatic Speech Recognition (ASR) systems.
The experimental setup involves a structured fine-tuning process with rigorous attention to dropout rates, training epochs, and batch sizes to enhance model performance.
[
Collection
]
[
|
...
]