<![CDATA[Voice based transaction input offers a practical alternative for reducing delayed and inaccurate financial recording in micro, small, and medium enterprises (MSMEs), particularly when owners record transactions while serving customers. This study empirically evaluates three Whisper automatic speech recognition variants for a voice driven financial transaction input system. The research workflow includes audio preparation, utterance segmentation, manual ground truth transcription, transaction field labeling, parameter efficient fine tuning using Low Rank Adaptation (LoRA), and hyperparameter tuning of learning rate and training epochs. The dataset comprises 28 transaction recordings from MSMEs in the Cilacap area, segmented into 140 short utterances containing local product names, transaction intents, and monetary values. Whisper tiny, Whisper base, and Whisper small are evaluated under zero shot and LoRA fine tuning settings using Word Error Rate (WER), Character Error Rate (CER), transaction field accuracy, and inference latency. The results show that Whisper small with a learning rate of 2e 4 and 20 epochs provides the best overall performance, achieving a WER of 5.86%, CER of 2.64%, field accuracy of 92.86%, and average latency of 1.36 seconds. Whisper base remains a practical alternative when lower latency is prioritized. These findings indicate that Whisper can support voice based transaction recording for MSMEs, particularly when nominal validation and user confirmation are incorporated before financial data are stored.]]>
Copyrights © 2026
