Comparing The Fine-Tuning and Performance of Whisper Pre-Trained Models for Turkish Speech Recognition Task

Despite the increasing success of Automatic Speech Recognition (ASR) technology, different results can be obtained for different languages. ASR technology has become more efficient due to the continuous development of artificial intelligence and machine learning algorithms. The development of many ready-made ASR systems has been made possible by this. The success rates of ready-made systems vary depending on the language used. The purpose of this study was to evaluate the performance of the Turkish ASR system created by finetuning the Whisper Small model and compare it to the existing Turkish ASR systems. Performance evaluations were conducted using the Word Error Rate (WER) metric. The Turkish ASR performance of the Whisper Small model was impressive on a wide vocabulary (WER %17.98). The fine-tuned model successfully recognized words and expressions belonging to different subject areas and produced accurate text predictions (WER %12.89). The Whisper Small ASR model has demonstrated the advantages of having a large vocabulary compared to the existing Turkish ASR systems. Compared to other systems, it produced more accurate results by working with a lower error rate.