DEEPSPEECH ARCHITECTURE BASED NON-NATIVE ENGLISH SPEAKER ASR USING FINE-TUNING METHOD

This research focuses on optimizing automatic speech recognition (ASR) for non- native English speakers, specifically Mandarin L1 speakers. The aim is to identify
the best dataset, hyper-parameters, and evaluation methods. The study uses
secondary datasets and evaluates using Word Error Rate (WER), Character Error
Rate (CER), and learning curves. The research applies fine-tuning techniques and
training from scratch to improve ASR performance. The fine-tuned model
achieved a WER of 0.45, CER of 0.19, and a loss of 35.29, outperforming the
model trained from scratch. The study found that a batch size of 8 for training and
2 for testing and validation resulted in the most efficient resource usage. The fine�tuned model performed best over 150 epochs, showing stable learning and
minimized loss.The results showed that fine-tuning outperformed training from
scratch, with lower WER and loss. The optimal GPU usage was achieved with
specific batch sizes, and the best models achieved a 0% WER/CER, ensuring
accurate speech transcription. Keyword: ASR, Non-native Speaker, Transfer Learning