Continuous Speech Recognition for the Lebanese Dialect: A Transfer Learning Approach

Abstract

Automatic Speech Recognition (ASR) has seen significant advancements over the past two decades, enabling systems to accurately transcribe and predict word sequences from input speech. Transfer Learning is the latest technology to be introduced and implemented to enhance ASR systems for low-resourced languages. This thesis introduces the first continuous, speaker-independent ASR model for Lebanese dialect, leveraging the power of Transfer Learning. After initial RNN-LSTM trials proved insufficient, and requiring the necessity of more robust technologies, the adoption of OpenAI’s Whisper model marked a significant shift in approach. By fine-tuning Whisper pre-trained model on Lebanese data, significant progress was achieved with a WER of 38% and a stable validation loss. This research presents the model design of the fine-tuned training process and also providing an analysis on the results produced, while contributing to the development of a small Lebanese dataset.