Research on speech enhancement algorithms in the airborne environment is of great significance to the security of airborne systems. Recently, the research focus of speech enhancement has turned from conventional unsupervised algorithms, like the log minimum mean square error estimator (log-MMSE), to the state-of-the-art masking-based long short-term memory (LSTM) method. However, each method has its characteristics and limitations, so they cannot always handle noise well. Besides, the requirements of clean speech and noise data for training a supervised speech enhancement model are difficult to satisfy in the real-world airborne environment. Therefore, in this work, to fully utilize the advantages of those two different methods without any data restrictions, we propose a novel adaptive gain mask (AGM) based teacher-student training approach for speech enhancement. In our method, the AGM, as a robust learning target for the student model, is devised by incorporating the estimated ideal ratio mask from the teacher model into the procedure of the log-MMSE approach. To get an appropriate tradeoff between the two methods, we adaptively update the AGM using a recursive weighting coefficient. Experiments on the real airborne data show that the proposed AGM-based method outperforms other baselines in terms of all essential objective metrics evaluated in this paper.
Airborne speech enhancement is always a major challenge for the security of airborne systems. Recently, multi-objective learning technology has become one of the mainstream methods of monaural speech enhancement. In this paper, we propose a novel multi-objective method for airborne speech enhancement, called the stacked multiscale densely connected temporal convolutional attention network (SMDTANet). More specifically, the core of SMDTANet includes three parts, namely a stacked multiscale feature extractor, a triple-attention-based temporal convolutional neural network (TA-TCNN), and a densely connected prediction module. The stacked multiscale feature extractor is leveraged to capture comprehensive feature information from noisy log-power spectra (LPS) inputs. Then, the TA-TCNN adopts a combination of these multiscale features and noisy amplitude modulation spectrogram (AMS) features as inputs to improve its powerful temporal modeling capability. In TA-TCNN, we integrate the advantages of channel attention, spatial attention, and T-F attention to design a novel triple-attention module, which can guide the network to suppress irrelevant information and emphasize informative features of different views. The densely connected prediction module is used to reliably control the flow of the information to provide an accurate estimation of clean LPS and the ideal ratio mask (IRM). Moreover, a new joint-weighted (JW) loss function is constructed to further improve the performance without adding to the model complexity. Extensive experiments on real-world airborne conditions show that our SMDTANet can obtain an on-par or better performance compared to other reference methods in terms of all the objective metrics of speech quality and intelligibility.
Research on speech enhancement algorithms in the airborne environment is of great significance to the security of airborne systems. Recently, many speech enhancement algorithms based on speech presence probability (SPP) have received considerable attention and discussion, like the traditional unbiased MMSE algorithm and the state-of-the-art masking-based deep learning approach. However, each method has its characteristics and limitations so that they cannot always handle noise well. In this work, to fully utilize the advantages of those two different methods, we propose a teacher-student training framework based on a new adaptive speech presence probability (ASPP) for speech enhancement. In our framework, the ASPP, as a robust learning target for the student model, is devised by incorporating the estimated ideal ratio mask (IRM) from the teacher model into the procedure of the unbiased MMSE approach. To get an appropriate tradeoff between the two methods, we adaptively update the ASPP using a recursive weighting coefficient, which depends upon the prior knowledge of the history frames, and the optimal boundary value of this weighting coefficient is obtained by independently varying the value of an adjustable scale factor. Experiments on the realistic airborne data set show that the proposed ASPP-based framework outperforms the other baselines in terms of all essential objective metrics evaluated in this paper. In addition, our framework using long short-term memory (LSTM) achieves the best speech enhancement performance with less computation.