Tao Li, Wenshuo Ge, Zhichao Wang, Zihao Cui, Yong Ma, Yingying Gao, Chao Deng, Shilei Zhang, Junlan Feng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We tend to assume most available speech corpora we use are either completely clean or completely noised. However, the reality is most of them are a mix of both. In this paper, we propose a semi-supervised speech enhancement framework to enhance such typical speech datasets. This framework includes an estimator to measure the speech purity. Utterances with high speech purity are considered clean, otherwise noised. For clean speech utterances, we follow the supervised learning mechanism to train a deep learning speech enhancement model. For noised speech, we update the model in an unsupervised manner. Hence, we design our training loss as a combination of the supervised loss and unsupervised loss. We refer to this framework as SemiEnhance. Experimental results show that SemiEnhance substantially improves the speech quality, and achieves new state-of-the-art results on benchmark datasets: 2022 DNS Challenge and NoiseX-92.

In this letter, a novel weighted mean square error (WMSE) is proposed to improve the DNN-based mask approximation method for speech enhancement, in which the weighting is closely related to the power exponent about noisy spectrum amplitude (NSA) base. The power exponents 0 and 2 separately reflect ideal amplitude masking (IAM) without any clippings and the indirect mapping (IM) on short-time spectral amplitude (STSA), and it is highly related to the enhanced spectrum and the performance of the enhanced signal based on the tests. Also, the experimental results show that the outstanding weighting is the noisy spectrum base with the power exponent 1 for the phase-unaware masking and results in better harmonic structure restoration. The objective function with the WMSE on the NSA (WMSE-NSA) can averagely improve 0.1 on the test of perceptual evaluation of speech quality (PESQ) and 1.7% on the test of short-time objective intelligibility (STOI) compared with the MSE-based mask approximation methods.

In this paper, a method for estimating the autoregressive parameters from a signal segment is proposed. The method is based on a deep neural network (DNN) in combination with the classical Levinson-Durbin recursion (LDR). The DNN acts as a pre-processor for the LDR and can be trained on different metrics commonly encountered in speech processing using a generalized analysis-by-synthesis (GABS) structure where the LDR acts as the encoder. Unlike end-to-end data-driven approaches, this structure ensures that the DNN is easy to train and initialize since the DNN only has to learn a simple mapping. The results confirm this and show that the proposed method produces an AR-spectrum that efficiently represents the speech spectrum in terms of the Itakura-Saito divergence, Kullback-Leibler divergence, log-spectral distortion, and speech distortion.

This paper proposes a linear prediction-based part-defined auto-encoder (PAE) network to enhance speech signal. The PAE is a defined decoder or a defined encoder network, based on efficient learning algorithm or classical model. In this paper, the PAE utilizes AR-Wiener filter as decoder part, and the AR-Wiener filter is modified as a linear prediction (LP) model by incorporating the modified factor from residual signal. The parameters of line spectral frequency (LSF) of speech and noise and the Wiener filtering mask are utilized for training targets. Finally, the proposed the LP-based PAE is compared with the baseline method, namely the Wiener filtering mask-based DNN. The PESQ and STOI results of the LP-based PAE are better than baseline method at lower signal noise ratio (SNR) levels.

The S-Transform which is used in many fields, is better in the methods of time-frequency analysis. Although the S-Transform improves the time-frequency analysis, it also increases the algorithm complexity, resulting in rapidly increasing the computing time and needing more memory. Therefore, the S-Transform is difficult to be applied for real-time procession of signal. Based on the characteristics that human being is insensitive to the voice frequency resolution, in this paper, we propose a compressing algorithm of the S-transform. Through re-sampling frequency, the algorithm decreases data size, the computing time and computer memory usage in the S-transform. The application of the algorithm in MFCC analysis shows the results is reliable under the condition of re-sampling frequency resolution .