Some recent studies have demonstrated the feasibility of single-stage neural text-to-speech, which does not need to generate mel-spectrograms but generates the raw waveforms directly from the text. Single-stage text-to-speech often faces two problems: a) the one-to-many mapping problem due to multiple speech variations and b) insufficiency of high frequency reconstruction due to the lack of supervision of ground-truth acoustic features during training. To solve the a) problem and generate more expressive speech, we propose a novel phoneme-level prosody modeling method based on a variational autoencoder with normalizing flows to model underlying prosodic information in speech. We also use the prosody predictor to support end-to-end expressive speech synthesis. Furthermore, we propose the dual parallel autoencoder to introduce supervision of the ground-truth acoustic features during training to solve the b) problem enabling our model to generate high-quality speech. We compare the synthesis quality with state-of-the-art text-to-speech systems on an internal expressive English dataset. Both qualitative and quantitative evaluations demonstrate the superiority and robustness of our method for lossless speech generation while also showing a strong capability in prosody modeling.
The accumulation of time series data and the absence of labels make time-series Anomaly Detection (AD) a self-supervised deep learning task. Single-assumption-based methods may only touch on a certain aspect of the whole normality, not sufficient to detect various anomalies. Among them, contrastive learning methods adopted for AD always choose negative pairs that are both normal to push away, which is objecting to AD tasks' purpose. Existing multi-assumption-based methods are usually two-staged, firstly applying a pre-training process whose target may differ from AD, so the performance is limited by the pre-trained representations. This paper proposes a deep Contrastive One-Class Anomaly detection method of time series (COCA), which combines the normality assumptions of contrastive learning and one-class classification. The key idea is to treat the representation and reconstructed representation as the positive pair of negative-samples-free contrastive learning, and we name it sequence contrast. Then we apply a contrastive one-class loss function composed of invariance and variance terms, the former optimizing loss of the two assumptions simultaneously, and the latter preventing hypersphere collapse. Extensive experiments conducted on four real-world time-series datasets show the superior performance of the proposed method achieves state-of-the-art. The code is publicly available at https://github.com/ruiking04/COCA.
It is critical to obtain high resolution features with long range dependency for dense prediction tasks such as semantic segmentation. To generate high-resolution output of size $H\times W$ from a low-resolution feature map of size $h\times w$ ($hw\ll HW$), a naive dense transformer incurs an intractable complexity of $\mathcal{O}(hwHW)$, limiting its application on high-resolution dense prediction. We propose a Dual-Flattening Transformer (DFlatFormer) to enable high-resolution output by reducing complexity to $\mathcal{O}(hw(H+W))$ that is multiple orders of magnitude smaller than the naive dense transformer. Decomposed queries are presented to retrieve row and column attentions tractably through separate transformers, and their outputs are combined to form a dense feature map at high resolution. To this end, the input sequence fed from an encoder is row-wise and column-wise flattened to align with decomposed queries by preserving their row and column structures, respectively. Row and column transformers also interact with each other to capture their mutual attentions with the spatial crossings between rows and columns. We also propose to perform attentions through efficient grouping and pooling to further reduce the model complexity. Extensive experiments on ADE20K and Cityscapes datasets demonstrate the superiority of the proposed dual-flattening transformer architecture with higher mIoUs.
With the rapid development of neural network architectures and speech processing models, singing voice synthesis with neural networks is becoming the cutting-edge technique of digital music production. In this work, in order to explore how to improve the quality and efficiency of singing voice synthesis, in this work, we use encoder-decoder neural models and a number of vocoders to achieve singing voice synthesis. We conduct experiments to demonstrate that the models can be trained using voice data with pitch information, lyrics and beat information, and the trained models can produce smooth, clear and natural singing voice that is close to real human voice. As the models work in the end-to-end manner, they allow users who are not domain experts to directly produce singing voice by arranging pitches, lyrics and beats.
With the rising of short video apps, such as TikTok, Snapchat and Kwai, advertisement in short-term user-generated videos (UGVs) has become a trending form of advertising. Prediction of user behavior without specific user profile is required by advertisers, as they expect to acquire advertisement performance in advance in the scenario of cold start. Current recommender system do not take raw videos as input; additionally, most previous work of Multi-Modal Machine Learning may not deal with unconstrained videos like UGVs. In this paper, we proposed a novel end-to-end self-organizing framework for user behavior prediction. Our model is able to learn the optimal topology of neural network architecture, as well as optimal weights, through training data. We evaluate our proposed method on our in-house dataset. The experimental results reveal that our model achieves the best performance in all our experiments.
Filter pruning has drawn more attention since resource constrained platform requires more compact model for deployment. However, current pruning methods suffer either from the inferior performance of one-shot methods, or the expensive time cost of iterative training methods. In this paper, we propose a balanced filter pruning method for both performance and pruning speed. Based on the filter importance criteria, our method is able to prune a layer with approximate layer-wise optimal pruning rate at preset loss variation. The network is pruned in the layer-wise way without the time consuming prune-retrain iteration. If a pre-defined pruning rate for the entire network is given, we also introduce a method to find the corresponding loss variation threshold with fast converging speed. Moreover, we propose the layer group pruning and channel selection mechanism for channel alignment in network with short connections. The proposed pruning method is widely applicable to common architectures and does not involve any additional training except the final fine-tuning. Comprehensive experiments show that our method outperforms many state-of-the-art approaches.
The applications of short-termuser generated video(UGV),such as snapchat, youtube short-term videos, booms recently,raising lots of multimodal machine learning tasks. Amongthem, learning the correspondence between audio and vi-sual information from videos is a challenging one. Mostprevious work of theaudio-visual correspondence(AVC)learning only investigated on constrained videos or simplesettings, which may not fit the application of UGV. In thispaper, we proposed new principles for AVC and introduced anew framework to set sight on the themes of videos to facili-tate AVC learning. We also released the KWAI-AD-AudViscorpus which contained 85432 short advertisement videos(around 913 hours) made by users. We evaluated our pro-posed approach on this corpus and it was able to outperformthe baseline by 23.15% absolute differenc
Named entity recognition (NER) and Relation extraction (RE) are two fundamental tasks in natural language processing applications. In practice, these two tasks are often to be solved simultaneously. Traditional multi-task learning models implicitly capture the correlations between NER and RE. However, there exist intrinsic connections between the output of NER and RE. In this study, we argue that an explicit interaction between the NER model and the RE model will better guide the training of both models. Based on the traditional multi-task learning framework, we design an interactive feature encoding method to capture the intrinsic connections between NER and RE tasks. In addition, we propose a recurrent interaction network to progressively capture the correlation between the two models. Empirical studies on two real-world datasets confirm the superiority of the proposed model.
Enabling bi-directional retrieval of images and texts is important for understanding the correspondence between vision and language. Existing methods leverage the attention mechanism to explore such correspondence in a fine-grained manner. However, most of them consider all semantics equally and thus align them uniformly, regardless of their diverse complexities. In fact, semantics are diverse (i.e. involving different kinds of semantic concepts), and humans usually follow a latent structure to combine them into understandable languages. It may be difficult to optimally capture such sophisticated correspondences in existing methods. In this paper, to address such a deficiency, we propose an Iterative Matching with Recurrent Attention Memory (IMRAM) method, in which correspondences between images and texts are captured with multiple steps of alignments. Specifically, we introduce an iterative matching scheme to explore such fine-grained correspondence progressively. A memory distillation unit is used to refine alignment knowledge from early steps to later ones. Experiment results on three benchmark datasets, i.e. Flickr8K, Flickr30K, and MS COCO, show that our IMRAM achieves state-of-the-art performance, well demonstrating its effectiveness. Experiments on a practical business advertisement dataset, named \Ads{}, further validates the applicability of our method in practical scenarios.
Facial appearance plays an important role in our social lives. Subjective perception of women's beauty depends on various face-related (e.g., skin, shape, hair) and environmental (e.g., makeup, lighting, angle) factors. Similar to cosmetic surgery in the physical world, virtual face beautification is an emerging field with many open issues to be addressed. Inspired by the latest advances in style-based synthesis and face beauty prediction, we propose a novel framework of face beautification. For a given reference face with a high beauty score, our GAN-based architecture is capable of translating an inquiry face into a sequence of beautified face images with referenced beauty style and targeted beauty score values. To achieve this objective, we propose to integrate both style-based beauty representation (extracted from the reference face) and beauty score prediction (trained on SCUT-FBP database) into the process of beautification. Unlike makeup transfer, our approach targets at many-to-many (instead of one-to-one) translation where multiple outputs can be defined by either different references or varying beauty scores. Extensive experimental results are reported to demonstrate the effectiveness and flexibility of the proposed face beautification framework.