Um estudo sobre a separação cega da fala em salas reverberantes

Abstract

Source separation is a fundamental step for the effective processing of speech signals in various applications, such as virtual assistants, videoconferencing systems, and hearing-assistance devices. In these contexts, the captured signals are often composed of mixtures of multiple speech sources and ambient noise, which makes recovering the original signals challenging. Signal separation methods are used either to filter out undesired speech signals or to recover them from mixtures. Separation tasks become particularly difficult when signals have been mixed in reverberant rooms due to the convolutive nature of the resulting mixtures. Techniques commonly used in such scenarios include Frequency-Domain Independent Component Analysis (FDICA) and Independent Vector Analysis (IVA), which rely on different optimization approaches such as natural gradient, fixed-point iteration (fast), and auxiliary functions. Factors such as the type of algorithm and the placement of microphones can affect separation performance in different acoustic scenarios—the former due to the adopted model and optimization strategy, and the latter due to the diversity of spatial information. This study aimed to investigate and compare these factors under specific experimental conditions. To this end, six blind speech separation methods for convolutive mixtures (three FDICA and three IVA) were applied to two-source mixtures recorded in four rooms with different reverberation times. The performance of the methods was then compared using the following metrics: Signal-to-Interference Ratio (SIR), Scale-Invariant Signal-to-Distortion Ratio (SI-SDR), and Perceptual Evaluation of Speech Quality (PESQ). In this stage, all factors except the method factor were blocked. Because SI-SDR and PESQ exhibited lower variability, they were also used to evaluate and compare, within each room, how microphone spacing (0 cm, 4 cm, and 57 cm) and the average distance between sources and microphones (75 cm, 140 cm, 210 cm) affected the performance of each of the three IVA methods selected from the initial analysis. Regarding algorithm comparison, SIR results were inconclusive. The SI-SDR and PESQ comparisons showed that fastFDICA was statistically inferior to the other methods. Moreover, IVAng outperformed the FDICA techniques and showed no statistically significant difference from the other IVA methods at a 5\% significance level.As for the evaluation of microphone placement, the results indicated that performance depended more on spatial information than on the specific algorithms. For each room, increasing the distance between microphones improved performance, although this effect diminished at higher reverberation times. Variations in the distances between microphones and sources did not produce statistically significant effects at any specific reverberation time.