Análisis comparativo de bibliotecas de Python para procesamiento de voz y reducción de ruido.

Abstract

Abstract: This Curriculum Integration Project addresses the comparative analysis of different algorithms focused on noise reduction in voice signals, implemented using Python libraries. The development was carried out using a dataset of 824 WAV audio files with a sampling frequency of 48 kHz. Four different libraries were evaluated: Wiener, NoiseReduce, Median, and Spectral Subtraction. For each method, evaluative metrics such as SNR, STOI, and PESQ were analyzed, in addition to computational performance based on processing speed in Mbps and execution times for each library. The results obtained indicate that the Wiener library demonstrates the best performance, followed closely by the Median library, while the NoiseReduce library shows the lowest performance. Furthermore, a low utilization of available hardware resources was observed, indicating that the observable differences are due to the internal workings of each library and not to a system limitation. This research paper presents a clear and structured methodology that allows for comparing the performance of different noise reduction libraries on speech signals.