Self-Supervised Learning (SSL) models have been successfully applied in various deep learning-based speech tasks, particularly those with a limited amount of data. However, the quality of SSL representations depends highly on the relatedness between the SSL training domain(s) and the target data domain. On the contrary, spectral feature (SF) extractors such as log Mel-filterbanks are hand-crafted non-learnable components, and could be more robust to domain shifts. The present work examines the assumption that combining non-learnable SF extractors to SSL models is an effective approach to low resource speech tasks. We propose a learnable and interpretable framework to combine SF and SSL representations. The proposed framework outperforms significantly both baseline and SSL models on Automatic Speech Recognition (ASR) and Speech Translation (ST) tasks on three low resource datasets. We additionally design a mixture of experts based combination model. This last model reveals that the relative contribution of SSL models over conventional SF extractors is very small in case of domain mismatch between SSL training set and the target language data.
In this work, we introduce LEAD, an approach to discover landmarks from an unannotated collection of category-specific images. Existing works in self-supervised landmark detection are based on learning dense (pixel-level) feature representations from an image, which are further used to learn landmarks in a semi-supervised manner. While there have been advances in self-supervised learning of image features for instance-level tasks like classification, these methods do not ensure dense equivariant representations. The property of equivariance is of interest for dense prediction tasks like landmark estimation. In this work, we introduce an approach to enhance the learning of dense equivariant representations in a self-supervised fashion. We follow a two-stage training approach: first, we train a network using the BYOL objective which operates at an instance level. The correspondences obtained through this network are further used to train a dense and compact representation of the image using a lightweight network. We show that having such a prior in the feature extractor helps in landmark detection, even under drastically limited number of annotations while also improving generalization across scale variations.
We propose a computational model of speech production combining a pre-trained neural articulatory synthesizer able to reproduce complex speech stimuli from a limited set of interpretable articulatory parameters, a DNN-based internal forward model predicting the sensory consequences of articulatory commands, and an internal inverse model based on a recurrent neural network recovering articulatory commands from the acoustic speech input. Both forward and inverse models are jointly trained in a self-supervised way from raw acoustic-only speech data from different speakers. The imitation simulations are evaluated objectively and subjectively and display quite encouraging performances.
Speech synthesis quality prediction has made remarkable progress with the development of supervised and self-supervised learning (SSL) MOS predictors but some aspects related to the data are still unclear and require further study. In this paper, we evaluate several MOS predictors based on wav2vec 2.0 and the NISQA speech quality prediction model to explore the role of the training data, the influence of the system type, and the role of cross-domain features in SSL models. Our evaluation is based on the VoiceMOS challenge dataset. Results show that SSL-based models show the highest correlation and lowest mean squared error compared to supervised models. The key point of this study is that benchmarking the statistical performance of MOS predictors alone is not sufficient to rank models since potential issues hidden in the data could bias the evaluated performances.
Machine learning, satellites or local sensors are key factors for a sustainable and resource-saving optimisation of agriculture and proved its values for the management of agricultural land. Up to now, the main focus was on the enlargement of data which were evaluated by means of supervised learning methods. Nevertheless, the need for labels is also a limiting and time-consuming factor, while in contrast, ongoing technological development is already providing an ever-increasing amount of unlabeled data. Self-supervised learning (SSL) could overcome this limitation and incorporate existing unlabeled data. Therefore, a crop type data set was utilized to conduct experiments with SSL and compare it to supervised methods. A unique feature of our data set from 2016 to 2018 was a divergent climatological condition in 2018 that reduced yields and affected the spectral fingerprint of the plants. Our experiments focused on predicting 2018 using SLL without or a few labels to clarify whether new labels should be collected for an unknown year. Despite these challenging conditions, the results showed that SSL contributed to higher accuracies. We believe that the results will encourage further improvements in the field of precision farming, why the SSL framework and data will be published (Marszalek, 2021).
Self-supervised learning of speech representations has achieved impressive results in improving automatic speech recognition (ASR). In this paper, we show that data selection is important for self-supervised learning. We propose a simple and effective unsupervised data selection method which selects acoustically similar speech to a target domain. It takes the discrete speech representation available in common self-supervised learning frameworks as input, and applies a contrastive data selection method on the discrete tokens. Through extensive empirical studies we show that our proposed method reduces the amount of required pre-training data and improves the downstream ASR performance. Pre-training on a selected subset of 6% of the general data pool results in 11.8% relative improvements in LibriSpeech test-other compared to pre-training on the full set. On Multilingual LibriSpeech French, German, and Spanish test sets, selecting 6% data for pre-training reduces word error rate by more than 15% relatively compared to the full set, and achieves competitive results compared to current state-of-the-art performances.
In many real world medical image classification settings we do not have access to samples of all possible disease classes, while a robust system is expected to give high performance in recognizing novel test data. We propose a generalized zero shot learning (GZSL) method that uses self supervised learning (SSL) for: 1) selecting anchor vectors of different disease classes; and 2) training a feature generator. Our approach does not require class attribute vectors which are available for natural images but not for medical images. SSL ensures that the anchor vectors are representative of each class. SSL is also used to generate synthetic features of unseen classes. Using a simpler architecture, our method matches a state of the art SSL based GZSL method for natural images and outperforms all methods for medical images. Our method is adaptable enough to accommodate class attribute vectors when they are available for natural images.
Medical image annotation typically requires expert knowledge and hence incurs time-consuming and expensive data annotation costs. To alleviate this burden, we propose a novel learning scenario, Exemplar Learning (EL), to explore automated learning processes for medical image segmentation with a single annotated image example. This innovative learning task is particularly suitable for medical image segmentation, where all categories of organs can be presented in one single image and annotated all at once. To address this challenging EL task, we propose an Exemplar Learning-based Synthesis Net (ELSNet) framework for medical image segmentation that enables innovative exemplar-based data synthesis, pixel-prototype based contrastive embedding learning, and pseudo-label based exploitation of the unlabeled data. Specifically, ELSNet introduces two new modules for image segmentation: an exemplar-guided synthesis module, which enriches and diversifies the training set by synthesizing annotated samples from the given exemplar, and a pixel-prototype based contrastive embedding module, which enhances the discriminative capacity of the base segmentation model via contrastive representation learning. Moreover, we deploy a two-stage process for segmentation model training, which exploits the unlabeled data with predicted pseudo segmentation labels. To evaluate this new learning framework, we conduct extensive experiments on several organ segmentation datasets and present an in-depth analysis. The empirical results show that the proposed exemplar learning framework produces effective segmentation results.
This paper describes our proposed integration system for the spoofing-aware speaker verification challenge. It consists of a robust spoofing-aware verification system that use the speaker verification and antispoofing embeddings extracted from specialized neural networks. First, an integration network, fed with the test utterance's speaker verification and spoofing embeddings, is used to compute a spoof-based score. This score is then linearly combined with the cosine similarity between the speaker verification embeddings from the enrollment and test utterances, thus obtaining the final scoring decision. Moreover, the integration network is trained using a one-class loss function to discriminate between target trials and unauthorized accesses. Our proposed system is evaluated in the ASVspoof19 database, exhibiting competitive performance compared to other integration approaches. In addition, we test, along with our integration approach, state of the art speaker verification and antispoofing systems based on self-supervised learning, yielding high-performance speech biometric systems.
Despite the success of Graph Neural Networks (GNNs) on various applications, GNNs encounter significant performance degradation when the amount of supervision signals, i.e., number of labeled nodes, is limited, which is expected as GNNs are trained solely based on the supervision obtained from the labeled nodes. On the other hand,recent self-supervised learning paradigm aims to train GNNs by solving pretext tasks that do not require any labeled nodes, and it has shown to even outperform GNNs trained with few labeled nodes. However, a major drawback of self-supervised methods is that they fall short of learning class discriminative node representations since no labeled information is utilized during training. To this end, we propose a novel semi-supervised method for graphs, GraFN, that leverages few labeled nodes to ensure nodes that belong to the same class to be grouped together, thereby achieving the best of both worlds of semi-supervised and self-supervised methods. Specifically, GraFN randomly samples support nodes from labeled nodes and anchor nodes from the entire graph. Then, it minimizes the difference between two predicted class distributions that are non-parametrically assigned by anchor-supports similarity from two differently augmented graphs. We experimentally show that GraFN surpasses both the semi-supervised and self-supervised methods in terms of node classification on real-world graphs. The source code for GraFN is available at https://github.com/Junseok0207/GraFN.
360{\deg} images are widely available over the last few years. This paper proposes a new technique for single 360{\deg} image depth prediction under open environments. Depth prediction from a 360{\deg} single image is not easy for two reasons. One is the limitation of supervision datasets - the currently available dataset is limited to indoor scenes. The other is the problems caused by Equirectangular Projection Format (ERP), commonly used for 360{\deg} images, that are coordinate and distortion. There is only one method existing that uses cube map projection to produce six perspective images and apply self-supervised learning using motion pictures for perspective depth prediction to deal with these problems. Different from the existing method, we directly use the ERP format. We propose a framework of direct use of ERP with coordinate conversion of correspondences and distortion-aware upsampling module to deal with the ERP related problems and extend a self-supervised learning method for open environments. For the experiments, we firstly built a dataset for the evaluation, and quantitatively evaluate the depth prediction in outdoor scenes. We show that it outperforms the state-of-the-art technique
Contrastive self-supervised learning has attracted significant research attention recently. It learns effective visual representations from unlabeled data by embedding augmented views of the same image close to each other while pushing away embeddings of different images. Despite its great success on ImageNet classification, COCO object detection, etc., its performance degrades on contrast-agnostic applications, e.g., medical image classification, where all images are visually similar to each other. This creates difficulties in optimizing the embedding space as the distance between images is rather small. To solve this issue, we present the first mix-up self-supervised learning framework for contrast-agnostic applications. We address the low variance across images based on cross-domain mix-up and build the pretext task based on two synergistic objectives: image reconstruction and transparency prediction. Experimental results on two benchmark datasets validate the effectiveness of our method, where an improvement of 2.5% ~ 7.4% in top-1 accuracy was obtained compared to existing self-supervised learning methods.
This work presents our end-to-end (E2E) automatic speech recognition (ASR) model targetting at robust speech recognition, called Integraded speech Recognition with enhanced speech Input for Self-supervised learning representation (IRIS). Compared with conventional E2E ASR models, the proposed E2E model integrates two important modules including a speech enhancement (SE) module and a self-supervised learning representation (SSLR) module. The SE module enhances the noisy speech. Then the SSLR module extracts features from enhanced speech to be used for speech recognition (ASR). To train the proposed model, we establish an efficient learning scheme. Evaluation results on the monaural CHiME-4 task show that the IRIS model achieves the best performance reported in the literature for the single-channel CHiME-4 benchmark (2.0% for the real development and 3.9% for the real test) thanks to the powerful pre-trained SSLR module and the fine-tuned SE module.
Unsupervised and self-supervised learning methods have leveraged unlabelled data to improve the pretrained models. However, these methods need significantly large amount of unlabelled data and the computational cost of training models with such large amount of data can be prohibitively high. We address this issue by using unlabelled data during finetuning, instead of pretraining. We propose acoustic model finetuning (FT) using labelled and unlabelled data. The model is jointly trained to learn representations to classify senones, as well as learn contextual acoustic representations. Our training objective is a combination of cross entropy loss, suitable for classification task, and contrastive loss, suitable to learn acoustic representations. The proposed approach outperforms conventional finetuning with 11.2% and 9.19% word error rate relative (WERR) reduction on Gujarati and Bengali languages respectively.
Non-coding RNA structure and function are essential to understanding various biological processes, such as cell signaling, gene expression, and post-transcriptional regulations. These are all among the core problems in the RNA field. With the rapid growth of sequencing technology, we have accumulated a massive amount of unannotated RNA sequences. On the other hand, expensive experimental observatory results in only limited numbers of annotated data and 3D structures. Hence, it is still challenging to design computational methods for predicting their structures and functions. The lack of annotated data and systematic study causes inferior performance. To resolve the issue, we propose a novel RNA foundation model (RNA-FM) to take advantage of all the 23 million non-coding RNA sequences through self-supervised learning. Within this approach, we discover that the pre-trained RNA-FM could infer sequential and evolutionary information of non-coding RNAs without using any labels. Furthermore, we demonstrate RNA-FM's effectiveness by applying it to the downstream secondary/3D structure prediction, SARS-CoV-2 genome structure and evolution prediction, protein-RNA binding preference modeling, and gene expression regulation modeling. The comprehensive experiments show that the proposed method improves the RNA structural and functional modelling results significantly and consistently. Despite only being trained with unlabelled data, RNA-FM can serve as the foundational model for the field.
Simplicial Embeddings (SEM) are representations learned through self-supervised learning (SSL), wherein a representation is projected into $L$ simplices of $V$ dimensions each using a softmax operation. This procedure conditions the representation onto a constrained space during pretraining and imparts an inductive bias for group sparsity. For downstream classification, we formally prove that the SEM representation leads to better generalization than an unnormalized representation. Furthermore, we empirically demonstrate that SSL methods trained with SEMs have improved generalization on natural image datasets such as CIFAR-100 and ImageNet. Finally, when used in a downstream classification task, we show that SEM features exhibit emergent semantic coherence where small groups of learned features are distinctly predictive of semantically-relevant classes.
User-defined keyword spotting is a task to detect new spoken terms defined by users. This can be viewed as a few-shot learning problem since it is unreasonable for users to define their desired keywords by providing many examples. To solve this problem, previous works try to incorporate self-supervised learning models or apply meta-learning algorithms. But it is unclear whether self-supervised learning and meta-learning are complementary and which combination of the two types of approaches is most effective for few-shot keyword discovery. In this work, we systematically study these questions by utilizing various self-supervised learning models and combining them with a wide variety of meta-learning algorithms. Our result shows that HuBERT combined with Matching network achieves the best result and is robust to the changes of few-shot examples.
Few-Shot classification aims at solving problems that only a few samples are available in the training process. Due to the lack of samples, researchers generally employ a set of training tasks from other domains to assist the target task, where the distribution between assistant tasks and the target task is usually different. To reduce the distribution gap, several lines of methods have been proposed, such as data augmentation and domain alignment. However, one common drawback of these algorithms is that they ignore the similarity task selection before training. The fundamental problem is to push the auxiliary tasks close to the target task. In this paper, we propose a novel task selecting algorithm, named Optimal Transport Task Selecting (OTTS), to construct a training set by selecting similar tasks for Few-Shot learning. Specifically, the OTTS measures the task similarity by calculating the optimal transport distance and completes the model training via a self-supervised strategy. By utilizing the selected tasks with OTTS, the training process of Few-Shot learning become more stable and effective. Other proposed methods including data augmentation and domain alignment can be used in the meantime with OTTS. We conduct extensive experiments on a variety of datasets, including MiniImageNet, CIFAR, CUB, Cars, and Places, to evaluate the effectiveness of OTTS. Experimental results validate that our OTTS outperforms the typical baselines, i.e., MAML, matchingnet, protonet, by a large margin (averagely 1.72\% accuracy improvement).
Self-supervised learning (SSL) to learn high-level speech representations has been a popular approach to building Automatic Speech Recognition (ASR) systems in low-resource settings. However, the common assumption made in literature is that a considerable amount of unlabeled data is available for the same domain or language that can be leveraged for SSL pre-training, which we acknowledge is not feasible in a real-world setting. In this paper, as part of the Interspeech Gram Vaani ASR challenge, we try to study the effect of domain, language, dataset size, and other aspects of our upstream pre-training SSL data on the final performance low-resource downstream ASR task. We also build on the continued pre-training paradigm to study the effect of prior knowledge possessed by models trained using SSL. Extensive experiments and studies reveal that the performance of ASR systems is susceptible to the data used for SSL pre-training. Their performance improves with an increase in similarity and volume of pre-training data. We believe our work will be helpful to the speech community in building better ASR systems in low-resource settings and steer research towards improving generalization in SSL-based pre-training for speech systems.
Speech representations learned from Self-supervised learning (SSL) models can benefit various speech processing tasks. However, utilizing SSL representations usually requires fine-tuning the pre-trained models or designing task-specific downstream models and loss functions, causing much memory usage and human labor. Recently, prompting in Natural Language Processing (NLP) has been found to be an efficient technique to leverage pre-trained language models (LMs). Specifically, prompt tuning optimizes a limited number of task-specific parameters with a fixed pre-trained model; as a result, only a small set of parameters is needed to be stored for each task. Prompt tuning improves computation and memory efficiency by leveraging the pre-trained LM's prediction ability. Nevertheless, such a paradigm is little studied in the speech community. We report in this paper the first exploration of the prompt tuning paradigm for speech processing tasks based on Generative Spoken Language Model (GSLM). Experiment results show that the prompt tuning technique achieves competitive performance in speech classification tasks with fewer trainable parameters than fine-tuning specialized downstream models. We further study the technique in challenging sequence generation tasks. Prompt tuning also demonstrates its potential, while the limitation and possible research directions are discussed in this paper. The source code is available on https://github.com/ga642381/SpeechPrompt.