Self-supervised training has shown promising gains in pretraining models and facilitating the downstream finetuning for speech recognition, like multilingual ASR. Most existing methods adopt a 2-stage scheme where the self-supervised loss is optimized in the first pretraining stage, and the standard supervised finetuning resumes in the second stage. In this paper, we propose an end-to-end (E2E) Joint Unsupervised and Supervised Training (JUST) method to combine the supervised RNN-T loss and the self-supervised contrastive and masked language modeling (MLM) losses. We validate its performance on the public dataset Multilingual LibriSpeech (MLS), which includes 8 languages and is extremely imbalanced. On MLS, we explore (1) JUST trained from scratch, and (2) JUST finetuned from a pretrained checkpoint. Experiments show that JUST can consistently outperform other existing state-of-the-art methods, and beat the monolingual baseline by a significant margin, demonstrating JUST's capability of handling low-resource languages in multilingual ASR. Our average WER of all languages outperforms average monolingual baseline by 33.3%, and the state-of-the-art 2-stage XLSR by 32%. On low-resource languages like Polish, our WER is less than half of the monolingual baseline and even beats the supervised transfer learning method which uses external supervision.
Recently, end-to-end (E2E) models, which allow to take spectral vector sequences of L2 (second-language) learners' utterances as input and produce the corresponding phone-level sequences as output, have attracted much research attention in developing mispronunciation detection (MD) systems. However, due to the lack of sufficient labeled speech data of L2 speakers for model estimation, E2E MD models are prone to overfitting in relation to conventional ones that are built on DNN-HMM acoustic models. To alleviate this critical issue, we in this paper propose two modeling strategies to enhance the discrimination capability of E2E MD models, each of which can implicitly leverage the phonetic and phonological traits encoded in a pretrained acoustic model and contained within reference transcripts of the training data, respectively. The first one is input augmentation, which aims to distill knowledge about phonetic discrimination from a DNN-HMM acoustic model. The second one is label augmentation, which manages to capture more phonological patterns from the transcripts of training data. A series of empirical experiments conducted on the L2-ARCTIC English dataset seem to confirm the efficacy of our E2E MD model when compared to some top-of-the-line E2E MD models and a classic pronunciation-scoring based method built on a DNN-HMM acoustic model.
The digital signal processing-based representations like the Mel-Frequency Cepstral Coefficient are well known to be a solid basis for various audio processing tasks. Alternatively, analog feature representations, relying on analog-electronics-feasible bandpass filtering, allow much lower system power consumption compared with the digital counterpart, while parity performance on traditional tasks like voice activity detection can be achieved. This work explores the possibility of using analog features on multiple speech processing tasks that vary in time dependencies: wake word detection, keyword spotting, and speaker identification. The results of this evaluation show that the analog features are still more power-efficient and competitive on simpler tasks than digital features but yield an increasing performance drop on more complex tasks when long-time correlations are present. We also introduce a novel theoretical framework based on information theory to understand this performance drop by quantifying information flow in feature calculation which helps identify the performance bottlenecks. The theoretical claims are experimentally validated, leading to a maximum of 6% increase of keyword spotting accuracy, even surpassing the digital baseline features. The proposed analog-feature-based systems could pave the way to achieving best-in-class accuracy and power consumption simultaneously.
The majority of Chinese characters are monophonic, i.e.their pronunciations are unique and thus can be induced easily using a check table. As for their counterparts, polyphonic characters have more than one pronunciation. To perform linguistic computation tasks related to spoken Mandarin Chinese, the correct pronunciation for each polyphone must be identified among several candidates according to its context. This process is called Polyphone Disambiguation, a key procedure in the Grapheme-to-phoneme (G2P) conversion step of a Chinese text-to-speech (TTS) system. The problem is well explored with both knowledge-based and learning-based approaches, yet it remains challenging due to the lack of publicly available datasets and complex language phenomenon concerned polyphone. In this paper, we propose a novel semi-supervised learning (SSL) framework for Mandarin Chinese polyphone disambiguation that can potentially leverage unlimited unlabeled text data. We explore the effect of various proxy labeling strategies including entropy-thresholding and lexicon-based labeling. As for the architecture, a pre-trained model of Electra is combined with Convolution BLSTM layers to fine-tune on our task. Qualitative and quantitative experiments demonstrate that our method achieves state-of-the-art performance in Mandarin Chinese polyphone disambiguation. In addition, we publish a novel dataset specifically for the polyphone disambiguation task to promote further researches.
We present a text-based tool for editing talking-head video that enables an iterative editing workflow. On each iteration users can edit the wording of the speech, further refine mouth motions if necessary to reduce artifacts and manipulate non-verbal aspects of the performance by inserting mouth gestures (e.g. a smile) or changing the overall performance style (e.g. energetic, mumble). Our tool requires only 2-3 minutes of the target actor video and it synthesizes the video for each iteration in about 40 seconds, allowing users to quickly explore many editing possibilities as they iterate. Our approach is based on two key ideas. (1) We develop a fast phoneme search algorithm that can quickly identify phoneme-level subsequences of the source repository video that best match a desired edit. This enables our fast iteration loop. (2) We leverage a large repository of video of a source actor and develop a new self-supervised neural retargeting technique for transferring the mouth motions of the source actor to the target actor. This allows us to work with relatively short target actor videos, making our approach applicable in many real-world editing scenarios. Finally, our refinement and performance controls give users the ability to further fine-tune the synthesized results.
In recent years, misinformation on the Web has become increasingly rampant. The research community has responded by proposing systems and challenges, which are beginning to be useful for (various subtasks of) detecting misinformation. However, most proposed systems are based on deep learning techniques which are fine-tuned to specific domains, are difficult to interpret and produce results which are not machine readable. This limits their applicability and adoption as they can only be used by a select expert audience in very specific settings. In this paper we propose an architecture based on a core concept of Credibility Reviews (CRs) that can be used to build networks of distributed bots that collaborate for misinformation detection. The CRs serve as building blocks to compose graphs of (i) web content, (ii) existing credibility signals --fact-checked claims and reputation reviews of websites--, and (iii) automatically computed reviews. We implement this architecture on top of lightweight extensions to Schema.org and services providing generic NLP tasks for semantic similarity and stance detection. Evaluations on existing datasets of social-media posts, fake news and political speeches demonstrates several advantages over existing systems: extensibility, domain-independence, composability, explainability and transparency via provenance. Furthermore, we obtain competitive results without requiring finetuning and establish a new state of the art on the Clef'18 CheckThat! Factuality task.
Generative adversarial networks (GANs) are increasingly attracting attention in the computer vision, natural language processing, speech synthesis and similar domains. However, evaluating the performance of GANs is still an open and challenging problem. Existing evaluation metrics primarily measure the dissimilarity between real and generated images using automated statistical methods. They often require large sample sizes for evaluation and do not directly reflect human perception of image quality. In this work, we introduce an evaluation metric called Neuroscore, for evaluating the performance of GANs, that more directly reflects psychoperceptual image quality through the utilization of brain signals. Our results show that Neuroscore has superior performance to the current evaluation metrics in that: (1) It is more consistent with human judgment; (2) The evaluation process needs much smaller numbers of samples; and (3) It is able to rank the quality of images on a per GAN basis. A convolutional neural network (CNN) based neuro-AI interface is proposed to predict Neuroscore from GAN-generated images directly without the need for neural responses. Importantly, we show that including neural responses during the training phase of the network can significantly improve the prediction capability of the proposed model. Codes and data can be referred at this link: https://github.com/villawang/Neuro-AI-Interface.
Assistive robots can potentially improve the quality of life and personal independence of elderly people by supporting everyday life activities. To guarantee a safe and intuitive interaction between human and robot, human intentions need to be recognized automatically. As humans communicate their intentions multimodally, the use of multiple modalities for intention recognition may not just increase the robustness against failure of individual modalities but especially reduce the uncertainty about the intention to be predicted. This is desirable as particularly in direct interaction between robots and potentially vulnerable humans a minimal uncertainty about the situation as well as knowledge about this actual uncertainty is necessary. Thus, in contrast to existing methods, in this work a new approach for multimodal intention recognition is introduced that focuses on uncertainty reduction through classifier fusion. For the four considered modalities speech, gestures, gaze directions and scene objects individual intention classifiers are trained, all of which output a probability distribution over all possible intentions. By combining these output distributions using the Bayesian method Independent Opinion Pool the uncertainty about the intention to be recognized can be decreased. The approach is evaluated in a collaborative human-robot interaction task with a 7-DoF robot arm. The results show that fused classifiers which combine multiple modalities outperform the respective individual base classifiers with respect to increased accuracy, robustness, and reduced uncertainty.
We present state-of-the-art automatic speech recognition (ASR) systems employing a standard hybrid DNN\/HMM architecture compared to an attention-based encoder-decoder design for the LibriSpeech task. Detailed descriptions of the system development, including model design, pretraining schemes, training schedules, and optimization approaches are provided for both system architectures. Both hybrid DNN/HMM and attention-based systems employ bi-directional LSTMs for acoustic modeling/encoding. For language modeling, we employ both LSTM and Transformer based architectures. All our systems are built using RWTHs open-source toolkits RASR and RETURNN. To the best knowledge of the authors, the results obtained when training on the full LibriSpeech training set, are the best published currently, both for the hybrid DNN/HMM and the attention-based systems. Our single hybrid system even outperforms previous results obtained from combining eight single systems. Our comparison shows that on the LibriSpeech 960h task, the hybrid DNN/HMM system outperforms the attention-based system by 15% relative on the clean and 40% relative on the other test sets in terms of word error rate. Moreover, experiments on a reduced 100h-subset of the LibriSpeech training corpus even show a more pronounced margin between the hybrid DNN/HMM and attention-based architectures.
Add to Chrome
Add to Firefox