Whispering is a special pronunciation style in which the vocal cords do not vibrate. Compared with voiced speech, whispering is noise-like because of the lack of a fundamental frequency. The energy of whispered speech is approximately 20 dB lower than that of voiced speech. Converting whispering into normal speech is an effective way to improve speech quality and/or intelligibility. In this paper, we propose a whisper-to-normal speech conversion method based on a sequence-to-sequence framework combined with an auditory attention mechanism. The proposed method does not require time aligning before conversion training, which makes it more applicable to real scenarios. In addition, the fundamental frequency is estimated from the mel frequency cepstral coefficient estimated by the proposed sequence-to-sequence framework. The voiced speech converted by the proposed method has appropriate length, which is determined adaptively by the proposed sequence-to-sequence model according to the source whispered speech. Experimental results show that the proposed sequence-to-sequence whisper-tonormal speech conversion method outperforms conventional DTW-based methods.

Although the complex spectrum-based speech enhancement (SE) methods have achieved significant performance, coupling amplitude and phase can lead to a compensation effect, where amplitude information is sacrificed to compensate for the phase that is harmful to SE. In addition, to further improve the performance of SE, many modules are stacked onto SE, resulting in increased model complexity that limits the application of SE. To address these problems, we proposed a dual-path network based on compressed frequency using Mamba. First, we extract amplitude and phase information through parallel dual branches. This approach leverages structured complex spectra to implicitly capture phase information and solves the compensation effect by decoupling amplitude and phase, and the network incorporates an interaction module to suppress unnecessary parts and recover missing components from the other branch. Second, to reduce network complexity, the network introduces a band-split strategy to compress the frequency dimension. To further reduce complexity while maintaining good performance, we designed a Mamba-based module that models the time and frequency dimensions under linear complexity. Finally, compared to baselines, our model achieves an average 8.3 times reduction in computational complexity while maintaining superior performance. Furthermore, it achieves a 25 times reduction in complexity compared to transformer-based models.

The brain-assisted target speaker extraction (TSE) aims to extract the attended speech from mixed speech by utilizing the brain neural activities, for example Electroencephalography (EEG). However, existing models overlook the issue of temporal misalignment between speech and EEG modalities, which hampers TSE performance. In addition, the speech encoder in current models typically uses basic temporal operations (e.g., one-dimensional convolution), which are unable to effectively extract target speaker information. To address these issues, this paper proposes a multi-scale and multi-modal alignment network (M3ANet) for brain-assisted TSE. Specifically, to eliminate the temporal inconsistency between EEG and speech modalities, the modal alignment module that uses a contrastive learning strategy is applied to align the temporal features of both modalities. Additionally, to fully extract speech information, multi-scale convolutions with GroupMamba modules are used as the speech encoder, which scans speech features at each scale from different directions, enabling the model to capture deep sequence information. Experimental results on three publicly available datasets show that the proposed model outperforms current state-of-the-art methods across various evaluation metrics, highlighting the effectiveness of our proposed method. The source code is available at: https://github.com/fchest/M3ANet.

Speech enhancement (SE) as a front-end task, especially in edge device deployment, requires low complexity and power consumption. However, deploying most artificial neural network (ANN)-based SE models while ensuring performance is challenging, particularly with the rapid development of large models in recent years. Spiking neural networks (SNNs) have shown potential in reducing power consumption. However, SNN-based SE models face two main challenges: on one hand, directly converting an ANN model to a SNN model introduces redundant modules, leading to difficulties in model training and information mismatch; on the other hand, the discrete binary activation and complex spatio-temporal dynamics of SNNs often result in information loss. To address these challenges, this paper proposes the SNN Speech Enhancement Network (SSE-Net). Unlike conventional approaches that convert ANNs to SNNs, all modules in SSE-Net are specifically designed for spike signal characteristics. Furthermore, the method innovatively develops a spiking feature extraction group that simultaneously converts information into spike signals while extracting original speech features, thereby effectively reducing information loss while significantly lowering both power consumption and computational complexity. The proposed method further designs information transformation blocks to convert spike signals back into continuous signals, fine-tuning and supplementing the speech information. Experimental results demonstrate that our proposed model achieves state of the art (SOTA) performance compared to the best SNN-SE models, with a 62% reduction in power proxy and 11% reduction in power cost, it also demonstrates strong competitiveness when compared with ANN-based SE models.