Strategies for preprocessing speech to enhance neural model efficiency in speech-to-text applications

A comprehensive study on the impact of advanced speech preprocessing strategies on the performance of speech-to-text models is presented. Our approach incorporates noise augmentation, speech rate descreasing, and anonymization to create a more robust training dataset. Additionally, we utilized voice cloning techniques to generate thousands of supplementary recordings, significantly expanding our dataset. These preprocessing strategies aim to improve the accuracy and efficiency of speech recognition systems. Our experiments demonstrate a notable reduction in Word Error Rate (WER) by an average of 3.5% and Character Error Rate (CER) by 4.5%. These results indicate that the applied preprocessing methods effectively enhance the model ability to generalize across diverse and challenging speech inputs. We provide insights into the preprocessing techniques, their implementation, and the subsequent impact on model performance, offering valuable guidance for future advancements in speech-to-text technologies. [The Polish National Center for Research and Development (NCBR) supported the project “ADMEDVOICE” (INFOSTRATEG4/0003/2022), focusing on voice processing applications in the healthcare sector.]