• Phonetics and Speech Sciences
• /
• v.15 no.4
• /
• pp.91-100
• /
• 2023

Reverberation time (T60) is a typical acoustic parameter that provides information about reverberation. Since the impacts of reverberation vary depending on the frequency bands even in the same space, frequency-dependent (FD) T60, which offers detailed insights into the acoustic environments, can be useful. However, most conventional blind T60 estimation methods, which estimate the T60 from speech signals, focus on fullband T60 estimation, and a few blind FDT60 estimation methods commonly show poor performance in the low-frequency bands. This paper introduces a modified approach based on Attentive pooling based Weighted Sum of Spectral Decay Rates (AWSSDR), previously proposed for blind T60 estimation, by extending its target from fullband T60 to FDT60. The experimental results show that the proposed method outperforms conventional blind FDT60 estimation methods on the acoustic characterization of environments (ACE) challenge evaluation dataset. Notably, it consistently exhibits excellent estimation performance in all frequency bands. This demonstrates that the mechanism of the AWSSDR method is valuable for blind FDT60 estimation because it reflects the FD variations in the impact of reverberation, aggregating information about FDT60 from the speech signal by processing the spectral decay rates associated with the physical properties of reverberation in each frequency band.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.497-504
• /
• 2020

Constructing speaker embeddings in speaker verification is an important issue. In general, a self-attention mechanism has been applied for speaker embedding encoding. Previous studies focused on training the self-attention in a high-level layer, such as the last pooling layer. In this case, the effect of low-level layers is not well represented in the speaker embedding encoding. In this study, we propose Masked Cross Self-Attentive Encoding (MCSAE) using ResNet. It focuses on training the features of both high-level and low-level layers. Based on multi-layer aggregation, the output features of each residual layer are used for the MCSAE. In the MCSAE, the interdependence of each input features is trained by cross self-attention module. A random masking regularization module is also applied to prevent overfitting problem. The MCSAE enhances the weight of frames representing the speaker information. Then, the output features are concatenated and encoded in the speaker embedding. Therefore, a more informative speaker embedding is encoded by using the MCSAE. The experimental results showed an equal error rate of 2.63 % using the VoxCeleb1 evaluation dataset. It improved performance compared with the previous self-attentive encoding and state-of-the-art methods.