Adversarial multi-task learning with inverse mapping for speech enhancement

Voice Activity Detection (VAD) is an important pre-processing step in a wide variety of speech processing systems. VAD should in a practical application be able to detect speech in both noisy and noise-free environments, while not introducing significant latency. In this work we propose using an adversarial multi-task learning method when training a supervised VAD. The method has been applied to the state-of-the-art VAD Waveform-based Voice Activity Detection. Additionally the performance of the VADis investigated under different algorithmic delays, which is an important factor in latency. Introducing adversarial multi-task learning to the model is observed to increase performance in terms of Area Under Curve (AUC), particularly in noisy environments, while the performance is not degraded at higher SNR levels. The adversarial multi-task learning is only applied in the training phase and thus introduces no additional cost in testing. Furthermore the correlation between performance and algorithmic delays is investigated, and it is observed that the VAD performance degradation is only moderate when lowering the algorithmic delay from 398 ms to 23 ms.

We propose a novel adversarial multi-task learning scheme, aiming at actively curtailing the inter-talker feature variability while maximizing its senone discriminability so as to enhance the performance of a deep neural network (DNN) based ASR system. We call the scheme speaker-invariant training (SIT). In SIT, a DNN acoustic model and a speaker classifier network are jointly optimized to minimize the senone (tied triphone state) classification loss, and simultaneously mini-maximize the speaker classification loss. A speaker-invariant and senone-discriminative deep feature is learned through this adversarial multi-task learning. With SIT, a canonical DNN acoustic model with significantly reduced variance in its output probabilities is learned with no explicit speaker-independent (SI) transformations or speaker-specific representations used in training or testing. Evaluated on the CHiME-3 dataset, the SIT achieves 4.99% relative word error rate (WER) improvement over the conventional SI acoustic model. With additional unsupervised speaker adaptation, the speaker-adapted (SA) SIT model achieves 4.86% relative WER gain over the SA SI acoustic model.

Dynamic controllable speech enhancement models based on quantile loss functions

Aggressive language detection (ALD), detecting the abusive and offensive language in texts, is one of the crucial applications in NLP community. Most existing works treat ALD as regular classification with neural models, while ignoring the inherent conflicts of social media text that they are quite unnormalized and irregular. In this work, we target improving the ALD by jointly performing text normalization (TN), via an adversarial multi-task learning framework. The private encoders for ALD and TN focus on the task-specific features retrieving, respectively, and the shared encoder learns the underlying common features over two tasks. During adversarial training, a task discriminator distinguishes the separate learning of ALD or TN. Experimental results on four ALD datasets show that our model outperforms all baselines under differing settings by large margins, demonstrating the necessity of joint learning the TN with ALD. Further analysis is conducted for a better understanding of our method.

Cross domains adversarial learning for Chinese named entity recognition for online medical consultation

This study aims to improve the source separation technique, DEMUCS, by revising the respective loss function. DEMUCS, developed by Facebook Team, is built on the Wave-U-Net and consists of convolutional layer encoding and decoding blocks with an LSTM layer in between. The applied loss function in DEMUCS contains wave-domain L1 distance and multi-scale short-time-Fourier-transform (STFT) loss.We present to revise the original loss by considering the perceptual metric scores, including perceptual speech quality (PESQ) and short-time objective intelligibility (STOI). The new loss function becomes a weighted sum of the original loss and the losses of STOI and PESQ, hoping to highlight the perceptual quality of the enhanced utterances.According to the preliminary experiments conducted on the VoiceBank-DEMUCS task, the DEMUCS network with the modified loss function provides the noise-corrupted utterances with superior objective perceptual metric scores (PESQ and STOI). These results indicate that the presented work benefits DEMUCS in speech enhancement performance.

Recently, deep neural networks (DNNs) have become the mainstream strategy for speech enhancement task because it can achieve the higher speech quality and intelligibility than the traditional methods. However, these DNN-based methods always need a large number of parallel corpus consisting of clean speech and noise to produce noisy data for the training of the DNN in order to improve the generalization of the network. As a result, this implies that many noisy speech signals that are collected in real environment cannot be used to train the DNN because of the lack of corresponding clean speech and noise. Additionally, as we know, noise varies with the time and scenario, so we cannot obtain parallel speech and noise due to infinite noise data and some limited speech data. Thus, the network training with unparallel speech and noise data is essential for the generalization of the network. To address this problem, we propose a novel parallel-data-free speech enhancement method, in which the cycle-consistent generative adversarial network (CycleGAN) and multi-objective learning are employed. Our method is also able to make best use of the benefits of multi-objective learning. On the training stage, we utilize two different encoders to encode the features of clean speech and noisy speech, respectively. Then, two forward generators are immediately used to predict the ideal time-frequency (T-F) mask and log-power spectrum (LPS) of clean speech. Two inverse generators are applied to map the magnitude spectrum (MS) and LPS of noisy speech, respectively. In addition, four discriminators are used to distinguish the real speech features from the generated features. Two encoders, four generators and four discriminators are simultaneously trained by using adversarial, identity-mapping, latent similarity and cycle-consistent loss. On the test stage, we directly utilize the forward generators and encoders to acquire the enhanced speech. The experimental results indicate that the proposed approach is able to achieve the better speech enhancement performance than the reference methods. Moreover, the proposed method is also effective to improve speech quality and intelligibility when the networks are trained under the parallel data.

Generative adversarial networks (GANs) have been increasingly used as feature mapping functions in speech enhancement, in which the noisy speech features are transformed to the clean ones through the generators. This article proposes a novel speech enhancement model based on a cycle-consistent relativistic GAN with Dilated Residual Networks and a Multi-attention mechanism. Using the adversarial loss, improved cycle-consistency losses, and an identity-mapping loss, a noisy-to-clean generator G and an inverse clean-to-noisy generator F simultaneously learn the forward and backward mappings between the source and target domains. To guarantee the stability of the training process, we replace vanilla GAN loss with relativistic average GAN loss and use spectral normalization in discriminators so that they conform to Lipschitz continuity. Furthermore, we employ two attention-based components as multi-attention mechanism to reduce importing signal distortion: attention U-net gates and dilated residual self-attention blocks. By employing these components, our proposed generators can capture long-term inner dependencies between elements of speech features and further preserve linguistic information. Experimental results on a public dataset indicate that the proposed model achieves state-of-the-art speech enhancement performance, especially in reducing speech distortion and improving signal overall quality. Compared with the representative GAN-based approaches, the proposed method significantly achieves the best performance in terms of STOI, CSIG, COVL, and CBAK objective metrics. Moreover, we demonstrate the contribution of each proposed component including relativistic average loss, attention U-net gate, self-attention layers, spectral normalization, and dilation operation by ten comparison systems.

In previous work, we proposed a variational autoencoder-based (VAE) Bayesian permutation training speech enhancement (SE) method (PVAE) which indicated that the SE performance of the traditional deep neural network-based (DNN) method could be improved by deep representation learning (DRL). Based on our previous work, we in this paper propose to use β-VAE to further improve PVAE's ability of representation learning. More specifically, our β-VAE can improve PVAE's capacity of disentangling different latent variables from the observed signal without the trade-off problem between disentanglement and signal reconstruction. This trade-off problem widely exists in previous β-VAE algorithms. Unlike the previous β-VAE algorithms, the proposed β-VAE strategy can also be used to optimize the DNN's structure. This means that the proposed method can not only improve PVAE's SE performance but also reduce the number of PVAE training parameters. The experimental results show that the proposed method can acquire better speech and noise latent representation than PVAE. Meanwhile, it also obtains a higher scale-invariant signal-to-distortion ratio, speech Quality, and speech intelligibility.

Detecting humor is a challenging task since words might share multiple valences and, depending on the context, the same words can be even used in offensive expressions. Neural network architectures based on Transformer obtain state-of-the-art results on several Natural Language Processing tasks, especially text classification. Adversarial learning, combined with other techniques such as multi-task learning, aids neural models learn the intrinsic properties of data. In this work, we describe our adversarial multi-task network, AMTL-Humor, used to detect and rate humor and offensive texts from Task 7 at SemEval-2021. Each branch from the model is focused on solving a related task, and consists of a BiLSTM layer followed by Capsule layers, on top of BERTweet used for generating contextualized embeddings. Our best model consists of an ensemble of all tested configurations, and achieves a 95.66% F1-score and 94.70% accuracy for Task 1a, while obtaining RMSE scores of 0.6200 and 0.5318 for Tasks 1b and 2, respectively.

The generative adversarial imitation learning (GAIL) has provided an adversarial learning framework for imitating expert policy from demonstrations in high-dimensional continuous tasks. However, almost all GAIL and its extensions only design a kind of reward function of logarithmic form in the adversarial training strategy with the Jensen-Shannon (JS) divergence for all complex environments. The fixed logarithmic type of reward function may be difficult to solve all complex tasks, and the vanishing gradients problem caused by the JS divergence will harm the adversarial learning process. In this paper, we propose a new algorithm named Wasserstein Distance guided Adversarial Imitation Learning (WDAIL) for promoting the performance of imitation learning (IL). There are three improvements in our method: (a) introducing the Wasserstein distance to obtain more appropriate measure in the adversarial training process, (b) using proximal policy optimization (PPO) in the reinforcement learning stage which is much simpler to implement and makes the algorithm more efficient, and (c) exploring different reward function shapes to suit different tasks for improving the performance. The experiment results show that the learning procedure remains remarkably stable, and achieves significant performance in the complex continuous control tasks of MuJoCo.

Deep Neural Network (DNN)-based mask estimation approach is an emerging algorithm in monaural speech enhancement. It is used to enhance speech signals from the noisy background by calculating either speech or noise dominant in a particular frame of the noisy speech signal. It can construct complex models for nonlinear processing. However, the limitation of the DNN-based mask algorithm is a generalization of the targeted population. Past research works focused on their target dataset because of time consumption for the audio recording session. Thus, in this work, different recording conditions were used to study the performance of the DNN-based mask estimation approach. The findings revealed that different language test dataset, as well as different conditions, may not give large impact in speech enhancement performance since the algorithm only learn the noise information. But, the performance of speech enhancement is promising when the trained model has been designed properly, especially given the less sample variations in the input dataset involved during the training session.

Most algorithms for speech enhancement in the spectral domain focus on the acquisition of an estimator of the clean speech parameter, such as spectrum or amplitude. Enhanced speech quality with these methods mainly depends upon the accuracy of the estimator. When the signal-to-noise ratio (SNR) becomes lower, e.g., less than 3 dB, the enhanced speech often shows unsatisfactory quality. Here, we propose a new method for speech enhancement in the spectral domain. It is based on our view that speech characteristics are perceived mainly by part of the spectral components which have higher local (or instantaneous) SNRs. These components have a special importance on speech SNR enhancement and speech understanding. Identifying these components in the noisy spectrum with a decision system and constructing enhanced speech just by the same noisy amplitude and phase, the result is similar to those popular algorithms. All that is needed is a local SNR decision for each spectral component. The threshold for the decision is derived and two popular methods for local SNR (LSNR) estimation are suggested. It shows better performance under lower SNR signals. Though the number of these kind of spectral components varies with different signal SNR, any algorithms which can pick up more than 90% of it is sufficient to match any existing methods. Primary considerations and results are shown.

To achieve high-quality voice communication technology without noise interference in flammable, explosive and strong electromagnetic environments, the speech enhancement technology of a fiber-optic external Fabry–Perot interferometric (EFPI) acoustic sensor based on deep learning is studied in this paper. The combination of a complex-valued convolutional neural network and a long short-term memory (CV-CNN-LSTM) model is proposed for speech enhancement in the EFPI acoustic sensing system. Moreover, the 3 × 3 coupler algorithm is used to demodulate voice signals. Then, the short-time Fourier transform (STFT) spectrogram features of voice signals are divided into a training set and a test set. The training set is input into the established CV-CNN-LSTM model for model training, and the test set is input into the trained model for testing. The experimental findings reveal that the proposed CV-CNN-LSTM model demonstrates exceptional speech enhancement performance, boasting an average Perceptual Evaluation of Speech Quality (PESQ) score of 3.148. In comparison to the CV-CNN and CV-LSTM models, this innovative model achieves a remarkable PESQ score improvement of 9.7% and 11.4%, respectively. Furthermore, the average Short-Time Objective Intelligibility (STOI) score witnesses significant enhancements of 4.04 and 2.83 when contrasted with the CV-CNN and CV-LSTM models, respectively.

The performance of deep neural network-based speech enhancement systems typically increases with the training dataset size. However, studies that investigated the effect of training dataset size on speech enhancement performance did not consider recent approaches, such as diffusion-based generative models. Diffusion models are typically trained with massive datasets for image generation tasks, but whether this is also required for speech enhancement is unknown. Moreover, studies that investigated the effect of training dataset size did not control for the data diversity. It is thus unclear whether the performance improvement was due to the increased dataset size or diversity. Therefore, we systematically investigate the effect of training dataset size on the performance of popular state-of-the-art discriminative and diffusion-based speech enhancement systems in matched conditions. We control for the data diversity by using a fixed set of speech utterances, noise segments and binaural room impulse responses to generate datasets of different sizes. We find that the diffusion-based systems perform the best relative to the discriminative systems in terms of objective metrics with datasets of 10 h or less. However, their objective metrics performance does not improve when increasing the training dataset size as much as the discriminative systems, and they are outperformed by the discriminative systems with datasets of 100 h or more.