Introducing Self-Supervised Learning Models for Spoken Query-Spoken Term Detection

The acoustic-to-word model based on the Connectionist Temporal Classification (CTC) criterion is a natural end-to-end (E2E) system directly targeting word as output unit. Two issues exist in the system: first, the current output of the CTC model relies on the current input and does not account for context weighted inputs. This is the hard alignment issue. Second, the word-based CTC model suffers from the out-of-vocabulary (OOV) issue. This means it can model only frequently occurring words while tagging the remaining words as OOV. Hence, such a model is limited in its capacity in recognizing only a fixed set of frequent words. In this study, we propose addressing these problems using a combination of attention mechanism and mixed-units. In particular, we introduce Attention CTC, Self-Attention CTC, Hybrid CTC, and Mixed-unit CTC. First, we blend attention modeling capabilities directly into the CTC network using Attention CTC and Self-Attention CTC. Second, to alleviate the OOV issue, we present Hybrid CTC which uses a word and letter CTC with shared hidden layers. The Hybrid CTC consults the letter CTC when the word CTC emits an OOV. Then, we propose a much better solution by training a Mixed-unit CTC which decomposes all the OOV words into sequences of frequent words and multi-letter units. Evaluated on a 3400 hours Microsoft Cortana voice assistant task, our final acoustic-to-word solution using attention and mixed-units achieves a relative reduction in word error rate (WER) over the vanilla word CTC by 12.09\%. Such an E2E model without using any language model (LM) or complex decoder also outperforms a traditional context-dependent (CD) phoneme CTC with strong LM and decoder by 6.79% relative.

Self-supervised learning (SSL) has shown promise in learning representations of audio that are useful for automatic speech recognition (ASR). But, training SSL models like wav2vec 2.0 requires a two-stage pipeline. In this paper we demonstrate a single-stage training of ASR models that can utilize both unlabeled and labeled data. During training, we alternately minimize two losses: an unsupervised masked Contrastive Predictive Coding (CPC) loss and the supervised audio-to-text alignment loss Connectionist Temporal Classification (CTC). We show that this joint training method directly optimizes performance for the downstream ASR task using unsupervised data while achieving similar word error rates to wav2vec 2.0 on the Librispeech 100-hours dataset. Finally, we postulate that solving the contrastive task is a regularization for the supervised CTC loss.

The availability of huge annotated datasets and deep learning have led to astounding Natural Language Processing (NLP) progress. However, there is a plethora of low-resource languages, lack the labelled corpora to train the models in the supervision. This kind of digital divide does not allow fair NLP development across language groups. Self-supervised learning (SSL) is a new paradigm in recent years that leverages large amounts of untagged text to learn powerful representations with little or no manual annotation. This paper is a review of self-supervised deep learning models in low-resource NLP tasks. It begins with the definition of the principles of the SSL and the difference between this approach and supervised and unsupervised approaches. We describe why methods such as masked language modeling, contrastive learning, and autoregressive modeling underlie modern pre-trained transformers such as BERT, GPT, and mBERT. A particular interest is paid to multilingual and cross lingual SSL schemes which allow knowledge transfer between high-resource and low resource languages. Low-resource tasks, including machine translation, sentiment analysis, speech-to-text, and information retrieval are reviewed. As per benchmark studies, even in small labelled samples, the SSL models can achieve big gains in accuracy and generalization. Computation cost, representational bias and morphologically rich and under-documented language evaluation, are however, problematic. The researchers observe that the study suffers certain limitations, such as over-reliance on high-resource pretraining information, inequity between linguistic groups, and the difficulty of deploying large-scale SSL models in resource-constrained circumstances. The future directions include lightweight multilingual models, federation of learning in NLP and symbolic linguistic knowledge and combination with SSL. Self-supervised deep learning bridges an essential gap between the high-resource and low-resource languages, and would be a highly valuable step toward inclusive, global NLP innovation.

End-to-end Automatic Speech Recognition (E2E ASR) significantly simplifies the training process of an ASR model. Connectionist Temporal Classification (CTC) is one of the most popular methods for E2E ASR training. Implicitly, CTC has a unique topology which is very useful for sequence modelling. However, we find that by changing to another topology, we can make it even more effective. In this paper, we propose a new CTC-like method, for E2E ASR training, by modifying the topology of original CTC, so that the well-known abuse of the blank label in CTC can be resolved theoretically. As we change the topology, a normalisation term is necessary, which makes the form of the final loss function similar to Maximum Mutual Information (MMI); we hence name our method MMI-CTC. In addition to maximising the posterior probability of the target sequence, the normalisation enables models to explicitly minimise the probability of competing hypothesis at the word sequence level. Our experimental results show that MMI-CTC is more efficient than CTC, and that the normalisation is essential for sequence training.

Recently, the acoustic-to-word model based on the Connectionist Temporal Classification (CTC) criterion was shown as a natural end-to-end model directly targeting words as output units. However, this type of word-based CTC model suffers from the out-of-vocabulary (OOV) issue as it can only model limited number of words in the output layer and maps all the remaining words into an OOV output node. Therefore, such word-based CTC model can only recognize the frequent words modeled by the network output nodes. It also cannot easily handle the hot-words which emerge after the model is trained. In this study, we improve the acoustic-to-word model with a hybrid CTC model which can predict both words and characters at the same time. With a shared-hidden-layer structure and modular design, the alignments of words generated from the word-based CTC and the character-based CTC are synchronized. Whenever the acoustic-to-word model emits an OOV token, we back off that OOV segment to the word output generated from the character-based CTC, hence solving the OOV or hot-words issue. Evaluated on a Microsoft Cortana voice assistant task, the proposed model can reduce the errors introduced by the OOV output token in the acoustic-to-word model by 30%.

Self-supervised learning (SSL) has achieved great success in speech-related tasks. While Transformer and Conformer architectures have dominated SSL backbones, encoders like Zipformer, which excel in automatic speech recognition (ASR), remain unexplored in SSL. Concurrently, inefficiencies in data processing within existing SSL training frameworks, such as fairseq, pose challenges in managing the growing volumes of training data. To address these issues, we propose k2SSL, an open-source framework that offers faster, more memory-efficient, and better-performing self-supervised speech representation learning, focusing on downstream ASR tasks. The optimized HuBERT and proposed Zipformer-based SSL systems exhibit substantial reductions in both training time and memory usage during SSL training. Experiments on LibriSpeech demonstrate that Zipformer Base significantly outperforms HuBERT and WavLM, achieving up to a 34.8% relative WER reduction compared to HuBERT Base after fine-tuning, along with a 3.5x pre-training speedup in GPU hours. When scaled to 60k hours of LibriLight data, Zipformer Large exhibits remarkable efficiency, matching HuBERT Large’s performance while requiring only 5/8 pre-training steps.

Connectionist Temporal Classification (CTC) and attention mechanism are two main approaches used in recent scene text recognition works. Compared with attention-based methods, CTC decoder has a much shorter inference time, yet a lower accuracy. To design an efficient and effective model, we propose the guided training of CTC (GTC), where CTC model learns a better alignment and feature representations from a more powerful attentional guidance. With the benefit of guided training, CTC model achieves robust and accurate prediction for both regular and irregular scene text while maintaining a fast inference speed. Moreover, to further leverage the potential of CTC decoder, a graph convolutional network (GCN) is proposed to learn the local correlations of extracted features. Extensive experiments on standard benchmarks demonstrate that our end-to-end model achieves a new state-of-the-art for regular and irregular scene text recognition and needs 6 times shorter inference time than attention-based methods.

Training deep learning models on medical images heavily depends on experts’ expensive and laborious manual labels. In addition, these images, labels, and even models themselves are not widely publicly accessible and suffer from various kinds of bias and imbalances. In this paper, chest X-ray pre-trained model via self-supervised contrastive learning (CheSS) was proposed to learn models with various representations in chest radiographs (CXRs). Our contribution is a publicly accessible pretrained model trained with a 4.8-M CXR dataset using self-supervised learning with a contrastive learning and its validation with various kinds of downstream tasks including classification on the 6-class diseases in internal dataset, diseases classification in CheXpert, bone suppression, and nodule generation. When compared to a scratch model, on the 6-class classification test dataset, we achieved 28.5% increase in accuracy. On the CheXpert dataset, we achieved 1.3% increase in mean area under the receiver operating characteristic curve on the full dataset and 11.4% increase only using 1% data in stress test manner. On bone suppression with perceptual loss, we achieved improvement in peak signal to noise ratio from 34.99 to 37.77, structural similarity index measure from 0.976 to 0.977, and root-square-mean error from 4.410 to 3.301 when compared to ImageNet pretrained model. Finally, on nodule generation, we achieved improvement in Fréchet inception distance from 24.06 to 17.07. Our study showed the decent transferability of CheSS weights. CheSS weights can help researchers overcome data imbalance, data shortage, and inaccessibility of medical image datasets. CheSS weight is available at https://github.com/mi2rl/CheSS.

Non-native speech causes automatic speech recognition systems to degrade in performance. Past strategies to address this challenge have considered model adaptation, accent classification with a model selection, alternate pronunciation lexicon, etc. In this study, we consider a recurrent neural network (RNN) with connectionist temporal classification (CTC) cost function trained on multi-accent English data including US (Native), Indian and Hispanic accents. We exploit dark knowledge from a model trained with the multi-accent data to train student models under the guidance of both a teacher model and CTC cost of target transcription. We show that transferring knowledge from a single RNN-CTC trained model toward a student model, yields better performance than the stand-alone teacher model. Since the outputs of different trained CTC models are not necessarily aligned, it is not possible to simply use an ensemble of CTC teacher models. To address this problem, we train accent specific models under the guidance of a single multi-accent teacher, which results in having multiple aligned and trained CTC models. Furthermore, we train a student model under the supervision of the accent-specific teachers, resulting in an even further complementary model, which achieves +20.1% relative Character Error Rate (CER) reduction compared to the baseline trained without any teacher. Having this effective multi-accent model, we can achieve further improvement for each accent by adapting the model to each accent. Using the accent specific model's outputs to regularize the adapting process (i.e., a knowledge distillation version of Kullback-Leibler (KL) divergence) results in even superior performance compared to the conventional approach using general teacher models.

Connectionist Temporal Classification (CTC) is an objective function for sequence learning and has shown promising results in speech and text recognition tasks. However, its inherent mechanism has not been investigated thoroughly. In this paper, we propose a theoretical explanation of CTC from the perspective of the Expectation-Maximization (EM) algorithm. Based on the EM analysis, we propose a pseudo-label-based L1 regularization and voting decoding algorithm to improve the performance of text recognition. The L1 regularization can reduce the pseudo-label estimation error, while the voting decoding algorithm modifies the built-in decoding logic of CTC and introduces a voting mechanism to the inference process. Experiments of handwritten text recognition show that the proposed method consistently improves over the CTC baseline and yields state-of-the-art results on three benchmark datasets.

End-to-end Frameworks with Connectionist Temporal Classification (CTC) have achieved great success in text recognition. Despite high accuracies with deep learning, CTC-based text recognition methods also suffer from poor alignment (character boundary positioning) in many applications. To address this issue, we propose an end-to-end text recognition method based on robust prototype learning. In the new CTC framework, we formulate the blank as the rejection of character classes and use the one-vs-all prototype classifier as the output layer of the convolutional neural network. For network learning, based on forced alignment between frames and character labels, the most aligned frame is up-weighted in CTC training strategy to reduce estimation errors in decoding. Experiments of handwritten text recognition on four benchmark datasets of different languages show that the proposed method consistently improves the accuracy and alignment of CTC-based text recognition baseline.

The performance of automatic speech recognition (ASR) may be degraded when accented speech is recognized because the speech has some linguistic differences from standard speech. Conventional accented speech recognition studies have utilized the accent embedding method, in which the accent embedding features are directly fed into the ASR network. Although the method improves the performance of accented speech recognition, it has some restrictions, such as increasing the computational costs. This study proposes an efficient method of training the ASR model for accented speech in a domain adversarial way based on the Domain Adversarial Neural Network (DANN). The DANN plays a role as a domain adaptation in which the training data and test data have different distributions. Thus, our approach is expected to construct a reliable ASR model for accented speech by reducing the distribution differences between accented speech and standard speech. DANN has three sub-networks: the feature extractor, the domain classifier, and the label predictor. To adjust the DANN for accented speech recognition, we constructed these three sub-networks independently, considering the characteristics of accented speech. In particular, we used an end-to-end framework based on Connectionist Temporal Classification (CTC) to develop the label predictor, a very important module that directly affects ASR results. To verify the efficiency of the proposed approach, we conducted several experiments of accented speech recognition for four English accents including Australian, Canadian, British (England), and Indian accents. The experimental results showed that the proposed DANN-based model outperformed the baseline model for all accents, indicating that the end-to-end domain adversarial training effectively reduced the distribution differences between accented speech and standard speech.

Voice interfaces are an essential part of contemporary human-computer interaction and speech-to-text (STT) systems allow one to work hands-free and enhance the accessibility. Nevertheless, the precision of voice command recognition is a major issue that is yet to be addressed particularly in noisy settings or in language with complex linguistic features. Current approaches can have issues with the consistency of real-time transcription and are not contextually aware and cannot work with limited or agglutinative languages like Hindi. In order to overcome these shortcomings, Enhancing Voice Command with Connectionist Temporal Classification (EVC-CTC) is proposed in this study. It is a strong framework which makes use of the CTC to enhance sequence matching between speech input and textual output. EVC-CTC combines both context modeling and language specific features to increase recognition accuracy particularly in the presence of real-world variability. The suggested approach is deployed as a modular and scalable STT pipeline that facilitates voice assistants that run on web-based platforms. Continuous voice streams are processed with EVC-CTC to preserve continuity of the context and dynamically respond to the background noise and variations in speakers. Experimental analysis indicates that EVC-CTC is much better than baseline models in Word Error Rate (WER) and semantic accuracy, especially in Hindi voice commands.

In this work, we investigate if the wav2vec 2.0 self-supervised pretraining helps mitigate the overfitting issues with connectionist temporal classification (CTC) training to reduce its performance gap with flat-start lattice-free MMI (E2E-LFMMI) for automatic speech recognition with limited training data. Towards that objective, we use the pretrained wav2vec 2.0 BASE model and fine-tune it on three different datasets including out-of-domain (Switchboard) and cross-lingual (Babel) scenarios. Our results show that for supervised adaptation of the wav2vec 2.0 model, both E2E-LFMMI and CTC achieve similar results; significantly outperforming the baselines trained only with supervised data. Fine-tuning the wav2vec 2.0 model with E2E-LFMMI and CTC we obtain the following relative WER improvements over the supervised baseline trained with E2E-LFMMI. We get relative improvements of 40% and 44% on the clean-set and 64% and 58% on the test set of Librispeech (100h) respectively. On Switchboard (300h) we obtain relative improvements of 33% and 35% respectively. Finally, for Babel languages, we obtain relative improvements of 26% and 23% on Swahili (38h) and 18% and 17% on Tagalog (84h) respectively.

End-to-end speech recognition models trained using joint Connectionist Temporal Classification (CTC)-Attention loss have gained popularity recently. In these models, a non-autoregressive CTC decoder is often used at inference time due to its speed and simplicity. However, such models are hard to personalize because of their conditional independence assumption that prevents output tokens from previous time steps to influence future predictions. To tackle this, we propose a novel two-way approach that first biases the encoder with attention over a predefined list of rare long-tail and out-of-vocabulary (OOV) words and then uses dynamic boosting and phone alignment network during decoding to further bias the subword pre-dictions. We evaluate our approach on open-source VoxPopuli and in-house medical datasets to showcase a 60% improvement in F1 score on domain-specific rare words over a strong CTC baseline.