End-to-end speaker diarization for an unknown number of speakers is addressed in this paper. Recently proposed end-to-end speaker diarization outperformed conventional clustering-based speaker diarization, but it has one drawback: it is less flexible in terms of the number of speakers. This paper proposes a method for encoder-decoder based attractor calculation (EDA), which first generates a flexible number of attractors from a speech embedding sequence. Then, the generated multiple attractors are multiplied by the speech embedding sequence to produce the same number of speaker activities. The speech embedding sequence is extracted using the conventional self-attentive end-to-end neural speaker diarization (SA-EEND) network. In a two-speaker condition, our method achieved a 2.69 % diarization error rate (DER) on simulated mixtures and a 8.07 % DER on the two-speaker subset of CALLHOME, while vanilla SA-EEND attained 4.56 % and 9.54 %, respectively. In unknown numbers of speakers conditions, our method attained a 15.29 % DER on CALLHOME, while the x-vector-based clustering method achieved a 19.43 % DER.
Inspired by recent developments in learning smoothed densities with empirical Bayes, we study variational autoencoders with a decoder that is tailored for the random variable $Y=X+N(0,\sigma^2 I_d)$. A notion of smoothed variational inference emerges where the smoothing is implicitly enforced by the noise model of the decoder; "implicit", since during training the encoder only sees clean samples. This is the concept of imaginary noise model, where the noise model dictates the functional form of the variational lower bound $\mathcal{L}(\sigma)$, but the noisy data are never seen during learning. The model is named $\sigma$-VAE. We prove that all $\sigma$-VAEs are equivalent to each other via a simple $\beta$-VAE expansion: $\mathcal{L}(\sigma_2) \equiv \mathcal{L}(\sigma_1,\beta)$, where $\beta=\sigma_2^2/\sigma_1^2$. We prove a similar result for the Laplace distribution in exponential families. Empirically, we report an intriguing power law $\mathcal{D}_{\rm KL} \sim \sigma^{-\nu}$ for the learned models and we study the inference in the $\sigma$-VAE for unseen noisy data. The experiments were performed on MNIST, where we show that quite remarkably the model can make reasonable inferences on extremely noisy samples even though it has not seen any during training. The vanilla VAE completely breaks down in this regime. We finish with a hypothesis (the XYZ hypothesis) on the findings here.
Following the rationale of end-to-end modeling, CTC, RNN-T or encoder-decoder-attention models for automatic speech recognition (ASR) use graphemes or grapheme-based subword units based on e.g. byte-pair encoding (BPE). The mapping from pronunciation to spelling is learned completely from data. In contrast to this, classical approaches to ASR employ secondary knowledge sources in the form of phoneme lists to define phonetic output labels and pronunciation lexica. In this work, we do a systematic comparison between grapheme- and phoneme-based output labels for an encoder-decoder-attention ASR model. We investigate the use of single phonemes as well as BPE-based phoneme groups as output labels of our model. To preserve a simplified and efficient decoder design, we also extend the phoneme set by auxiliary units to be able to distinguish homophones. Experiments performed on the Switchboard 300h and LibriSpeech benchmarks show that phoneme-based modeling is competitive to grapheme-based encoder-decoder-attention modeling.
In sequence-to-sequence learning, e.g., natural language generation, the decoder relies on the attention mechanism to efficiently extract information from the encoder. While it is common practice to draw information from only the last encoder layer, recent work has proposed to use representations from different encoder layers for diversified levels of information. Nonetheless, the decoder still obtains only a single view of the source sequences, which might lead to insufficient training of the encoder layer stack due to the hierarchy bypassing problem. In this work, we propose layer-wise multi-view decoding, where for each decoder layer, together with the representations from the last encoder layer, which serve as a global view, those from other encoder layers are supplemented for a stereoscopic view of the source sequences. Systematic experiments and analyses show that we successfully address the hierarchy bypassing problem, require almost negligible parameter increase, and substantially improve the performance of sequence-to-sequence learning with deep representations on five diverse tasks, i.e., machine translation, abstractive summarization, image captioning, video captioning, medical report generation, and paraphrase generation. In particular, our approach achieves new state-of-the-art results on ten benchmark datasets, including a low-resource machine translation dataset and two low-resource medical report generation datasets.
In software development through integrated development environments (IDEs), code completion is one of the most widely used features. Nevertheless, majority of integrated development environments only support completion of methods and APIs, or arguments. In this paper, we introduce IntelliCode Compose $-$ a general-purpose multilingual code completion tool which is capable of predicting sequences of code tokens of arbitrary types, generating up to entire lines of syntactically correct code. It leverages state-of-the-art generative transformer model trained on 1.2 billion lines of source code in Python, $C\#$, JavaScript and TypeScript programming languages. IntelliCode Compose is deployed as a cloud-based web service. It makes use of client-side tree-based caching, efficient parallel implementation of the beam search decoder, and compute graph optimizations to meet edit-time completion suggestion requirements in the Visual Studio Code IDE and Azure Notebook. Our best model yields an average edit similarity of $86.7\%$ and a perplexity of 1.82 for Python programming language.
Recently, end-to-end multi-speaker text-to-speech (TTS) systems gain success in the situation where a lot of high-quality speech plus their corresponding transcriptions are available. However, laborious paired data collection processes prevent many institutes from building multi-speaker TTS systems of great performance. In this work, we propose a semi-supervised learning approach for multi-speaker TTS. A multi-speaker TTS model can learn from the untranscribed audio via the proposed encoder-decoder framework with discrete speech representation. The experiment results demonstrate that with only an hour of paired speech data, no matter the paired data is from multiple speakers or a single speaker, the proposed model can generate intelligible speech in different voices. We found the model can benefit from the proposed semi-supervised learning approach even when part of the unpaired speech data is noisy. In addition, our analysis reveals that different speaker characteristics of the paired data have an impact on the effectiveness of semi-supervised TTS.
Recent advances in end-to-end unsupervised learning has significantly improved the performance of monocular depth prediction and alleviated the requirement of ground truth depth. Although a plethora of work has been done in enforcing various structural constraints by incorporating multiple losses utilising smoothness, left-right consistency, regularisation and matching surface normals, a few of them take into consideration multi-scale structures present in real world images. Most works utilise a VGG16 or ResNet50 model pre-trained on ImageNet weights for predicting depth. We propose a deep feature fusion method utilising features at multiple scales for learning self-supervised depth from scratch. Our fusion network selects features from both upper and lower levels at every level in the encoder network, thereby creating multiple feature pyramid sub-networks that are fed to the decoder after applying the CoordConv solution. We also propose a refinement module learning higher scale residual depth from a combination of higher level deep features and lower level residual depth using a pixel shuffling framework that super-resolves lower level residual depth. We select the KITTI dataset for evaluation and show that our proposed architecture can produce better or comparable results in depth prediction.
Many semi- and weakly-supervised approaches have been investigated for overcoming the labeling cost of building high quality speech recognition systems. On the challenging task of transcribing social media videos in low-resource conditions, we conduct a large scale systematic comparison between two self-labeling methods on one hand, and weakly-supervised pretraining using contextual metadata on the other. We investigate distillation methods at the frame level and the sequence level for hybrid, encoder-only CTC-based, and encoder-decoder speech recognition systems on Dutch and Romanian languages using 27,000 and 58,000 hours of unlabeled audio respectively. Although all approaches improved upon their respective baseline WERs by more than 8%, sequence-level distillation for encoder-decoder models provided the largest relative WER reduction of 20% compared to the strongest data-augmented supervised baseline.
Recent studies reveal the potential of recurrent neural network transducer (RNN-T) for end-to-end (E2E) speech recognition. Among some most popular E2E systems including RNN-T, Attention Encoder-Decoder (AED), and Connectionist Temporal Classification (CTC), RNN-T has some clear advantages given that it supports streaming recognition and does not have frame-independency assumption. Although significant progresses have been made for RNN-T research, it is still facing performance challenges in terms of training speed and accuracy. We propose attention-based transducer with modification over RNN-T in two aspects. First, we introduce chunk-wise attention in the joint network. Second, self-attention is introduced in the encoder. Our proposed model outperforms RNN-T for both training speed and accuracy. For training, we achieves over 1.7x speedup. With 500 hours LAIX non-native English training data, attention-based transducer yields ~10.6% WER reduction over baseline RNN-T. Trained with full set of over 10K hours data, our final system achieves ~5.5% WER reduction over that trained with the best Kaldi TDNN-f recipe. After 8-bit weight quantization without WER degradation, RTF and latency drop to 0.34~0.36 and 268~409 milliseconds respectively on a single CPU core of a production server.
We propose a framework called HyperVAE for encoding distributions of distributions. When a target distribution is modeled by a VAE, its neural network parameters \theta is drawn from a distribution p(\theta) which is modeled by a hyper-level VAE. We propose a variational inference using Gaussian mixture models to implicitly encode the parameters \theta into a low dimensional Gaussian distribution. Given a target distribution, we predict the posterior distribution of the latent code, then use a matrix-network decoder to generate a posterior distribution q(\theta). HyperVAE can encode the parameters \theta in full in contrast to common hyper-networks practices, which generate only the scale and bias vectors as target-network parameters. Thus HyperVAE preserves much more information about the model for each task in the latent space. We discuss HyperVAE using the minimum description length (MDL) principle and show that it helps HyperVAE to generalize. We evaluate HyperVAE in density estimation tasks, outlier detection and discovery of novel design classes, demonstrating its efficacy.
Always-on spoken language interfaces, e.g. personal digital assistants, rely on a wake word to start processing spoken input. We present novel methods to train a hybrid DNN/HMM wake word detection system from partially labeled training data, and to use it in on-line applications: (i) we remove the prerequisite of frame-level alignments in the LF-MMI training algorithm, permitting the use of un-transcribed training examples that are annotated only for the presence/absence of the wake word; (ii) we show that the classical keyword/filler model must be supplemented with an explicit non-speech (silence) model for good performance; (iii) we present an FST-based decoder to perform online detection. We evaluate our methods on two real data sets, showing 50%--90% reduction in false rejection rates at pre-specified false alarm rates over the best previously published figures, and re-validate them on a third (large) data set.
Can artificial neural networks learn to represent inflectional morphology and generalize to new words as human speakers do? Kirov and Cotterell (2018) argue that the answer is yes: modern Encoder-Decoder (ED) architectures learn human-like behavior when inflecting English verbs, such as extending the regular past tense form -(e)d to novel words. However, their work does not address the criticism raised by Marcus et al. (1995): that neural models may learn to extend not the regular, but the most frequent class -- and thus fail on tasks like German number inflection, where infrequent suffixes like -s can still be productively generalized. To investigate this question, we first collect a new dataset from German speakers (production and ratings of plural forms for novel nouns) that is designed to avoid sources of information unavailable to the ED model. The speaker data show high variability, and two suffixes evince 'regular' behavior, appearing more often with phonologically atypical inputs. Encoder-decoder models do generalize the most frequently produced plural class, but do not show human-like variability or 'regular' extension of these other plural markers. We conclude that modern neural models may still struggle with minority-class generalization.
Depth completion recovers a dense depth map from sensor measurements. Current methods are mostly tailored for very sparse depth measurements from LiDARs in outdoor settings, while for indoor scenes Time-of-Flight (ToF) or structured light sensors are mostly used. These sensors provide semi-dense maps, with dense measurements in some regions and almost empty in others. We propose a new model that takes into account the statistical difference between such regions. Our main contribution is a new decoder modulation branch added to the encoder-decoder architecture. The encoder extracts features from the concatenated RGB image and raw depth. Given the mask of missing values as input, the proposed modulation branch controls the decoding of a dense depth map from these features differently for different regions. This is implemented by modifying the spatial distribution of output signals inside the decoder via Spatially-Adaptive Denormalization (SPADE) blocks. Our second contribution is a novel training strategy that allows us to train on a semi-dense sensor data when the ground truth depth map is not available. Our model achieves the state of the art results on indoor Matterport3D dataset. Being designed for semi-dense input depth, our model is still competitive with LiDAR-oriented approaches on the KITTI dataset. Our training strategy significantly improves prediction quality with no dense ground truth available, as validated on the NYUv2 dataset.
Emotional voice conversion aims to convert the emotion of speech from one state to another while preserving the linguistic content and speaker identity. The prior studies on emotional voice conversion are mostly carried out under the assumption that emotion is speaker-dependent. We consider that there is a common code between speakers for emotional expression in a spoken language, therefore, a speaker-independent mapping between emotional states is possible. In this paper, we propose a speaker-independent emotional voice conversion framework, that can convert anyone's emotion without the need for parallel data. We propose a VAW-GAN based encoder-decoder structure to learn the spectrum and prosody mapping. We perform prosody conversion by using continuous wavelet transform (CWT) to model the temporal dependencies. We also investigate the use of F0 as an additional input to the decoder to improve emotion conversion performance. Experiments show that the proposed speaker-independent framework achieves competitive results for both seen and unseen speakers.
Discourse representation tree structure (DRTS) parsing is a novel semantic parsing task which has been concerned most recently. State-of-the-art performance can be achieved by a neural sequence-to-sequence model, treating the tree construction as an incremental sequence generation problem. Structural information such as input syntax and the intermediate skeleton of the partial output has been ignored in the model, which could be potentially useful for the DRTS parsing. In this work, we propose a structural-aware model at both the encoder and decoder phase to integrate the structural information, where graph attention network (GAT) is exploited for effectively modeling. Experimental results on a benchmark dataset show that our proposed model is effective and can obtain the best performance in the literature.
Deep Neural Networks are increasingly used in video frame interpolation tasks such as frame rate changes as well as generating fake face videos. Our project aims to apply recent advances in Deep video interpolation to increase the temporal resolution of fluorescent microscopy time-lapse movies. To our knowledge, there is no previous work that uses Convolutional Neural Networks (CNN) to generate frames between two consecutive microscopy images. We propose a fully convolutional autoencoder network that takes as input two images and generates upto seven intermediate images. Our architecture has two encoders each with a skip connection to a single decoder. We evaluate the performance of several variants of our model that differ in network architecture and loss function. Our best model out-performs state of the art video frame interpolation algorithms. We also show qualitative and quantitative comparisons with state-of-the-art video frame interpolation algorithms. We believe deep video interpolation represents a new approach to improve the time-resolution of fluorescent microscopy.
Machine reading comprehension (MRC) aims to teach machines to read and comprehend human languages, which is a long-standing goal of natural language processing (NLP). With the burst of deep neural networks and the evolution of contextualized language models (CLMs), the research of MRC has experienced two significant breakthroughs. MRC and CLM, as a phenomenon, have a great impact on the NLP community. In this survey, we provide a comprehensive and comparative review on MRC covering overall research topics about 1) the origin and development of MRC and CLM, with a particular focus on the role of CLMs; 2) the impact of MRC and CLM to the NLP community; 3) the definition, datasets, and evaluation of MRC; 4) general MRC architecture and technical methods in the view of two-stage Encoder-Decoder solving architecture from the insights of the cognitive process of humans; 5) previous highlights, emerging topics, and our empirical analysis, among which we especially focus on what works in different periods of MRC researches. We propose a full-view categorization and new taxonomies on these topics. The primary views we have arrived at are that 1) MRC boosts the progress from language processing to understanding; 2) the rapid improvement of MRC systems greatly benefits from the development of CLMs; 3) the theme of MRC is gradually moving from shallow text matching to cognitive reasoning.
Past few years have witnessed the artificial intelligence inspired evolution in various medical fields. The diagnosis and treatment of gliomas -- one of the most commonly seen brain tumors with low survival rate -- rely heavily on the computer assisted segmentation process undertaken on the magnetic resonance imaging (MRI) scans. Although the encoder-decoder shaped deep learning networks have been the de facto standard style for semantic segmentation tasks in medical imaging analysis, enormous effort is still required to be spent on designing the detailed architecture of the down-sampling and up-sampling blocks. In this work, we propose a neural architecture search (NAS) based solution to brain tumor segmentation tasks on multimodal volumetric MRI scans. Three sets of candidate operations are composed respectively for three kinds of basic building blocks in which each operation is assigned with a specific probabilistic parameter to be learned. Through alternately updating the weights of operations and the other parameters in the network, the searching mechanism ends up with two optimal structures for the upward and downward blocks. Moreover, the developed solution also integrates normalization and patching strategies tailored for brain MRI processing. Extensive comparative experiments on the BraTS 2019 dataset demonstrate that the proposed algorithm not only could relieve the pressure of fabricating block architectures but also possesses competitive feasibility and scalability.
Infrared and visible image fusion (IVIF) expects to obtain images that retain thermal radiation information from infrared images and texture details from visible images. In this paper, a model-based convolutional neural network (CNN) model, referred to as Algorithm Unrolling Image Fusion (AUIF), is proposed to overcome the shortcomings of traditional CNN-based IVIF models. The proposed AUIF model starts with the iterative formulas of two traditional optimization models, which are established to accomplish two-scale decomposition, i.e., separating low-frequency base information and high-frequency detail information from source images. Then the algorithm unrolling is implemented where each iteration is mapped to a CNN layer and each optimization model is transformed into a trainable neural network. Compared with the general network architectures, the proposed framework combines the model-based prior information and is designed more reasonably. After the unrolling operation, our model contains two decomposers (encoders) and an additional reconstructor (decoder). In the training phase, this network is trained to reconstruct the input image. While in the test phase, the base (or detail) decomposed feature maps of infrared/visible images are merged respectively by an extra fusion layer, and then the decoder outputs the fusion image. Qualitative and quantitative comparisons demonstrate the superiority of our model, which can robustly generate fusion images containing highlight targets and legible details, exceeding the state-of-the-art methods. Furthermore, our network has fewer weights and faster speed.
This paper describes a methodology to produce a 7-classes land cover map of urban areas from very high resolution images and limited noisy labeled data. The objective is to make a segmentation map of a large area (a french department) with the following classes: asphalt, bare soil, building, grassland, mineral material (permeable artificialized areas), forest and water from 20cm aerial images and Digital Height Model. We created a training dataset on a few areas of interest aggregating databases, semi-automatic classification, and manual annotation to get a complete ground truth in each class. A comparative study of different encoder-decoder architectures (U-Net, U-Net with Resnet encoders, Deeplab v3+) is presented with different loss functions. The final product is a highly valuable land cover map computed from model predictions stitched together, binarized, and refined before vectorization.