This model combines transcriptions from multiple sources separated by '/' to generate an optimal transcription. It is fine-tuned on a dataset where each sample has three candidate transcriptions and a reference transcription.

Model trained on 21000 samples for 10 epochs with T5-small as the base model.

Training Loss: 0.008486982434988022

Test Loss: 0.012162444764789124 Word Error Rate (WER): 0.1033691040678812