Automatic Speech Recognition (ASR) systems have found their use in numerous industrial applications in very diverse domains. Since domain-specific systems perform better than their generic counterparts on in-domain evaluation, the need for memory and compute-efficient domain adaptation is obvious. Particularly, adapting parameter-heavy transformer-based language models used for rescoring ASR hypothesis is challenging. In this work, we overcome the problem using prompt-tuning, a methodology that trains a small number of domain token embedding parameters to prime a transformer-based LM to a particular domain. With just a handful of extra parameters per domain, we achieve much better perplexity scores over the baseline of using an unadapted LM. Despite being parameter-efficient, these improvements are comparable to those of fully-fine-tuned models with hundreds of millions of parameters. We replicate our findings in perplexity numbers to Word Error Rate in a domain-specific ASR system for one such domain.
Speech summarization is typically performed by using a cascade of speech recognition and text summarization models. End-to-end modeling of speech summarization models is challenging due to memory and compute constraints arising from long input audio sequences. Recent work in document summarization has inspired methods to reduce the complexity of self-attentions, which enables transformer models to handle long sequences. In this work, we introduce a single model optimized end-to-end for speech summarization. We apply the restricted self-attention technique from text-based models to speech models to address the memory and compute constraints. We demonstrate that the proposed model learns to directly summarize speech for the How-2 corpus of instructional videos. The proposed end-to-end model outperforms the previously proposed cascaded model by 3 points absolute on ROUGE. Further, we consider the spoken language understanding task of predicting concepts from speech inputs and show that the proposed end-to-end model outperforms the cascade model by 4 points absolute F-1.
We propose a spatial diffuseness feature for deep neural network (DNN)-based automatic speech recognition to improve recognition accuracy in reverberant and noisy environments. The feature is computed in real-time from multiple microphone signals without requiring knowledge or estimation of the direction of arrival, and represents the relative amount of diffuse noise in each time and frequency bin. It is shown that using the diffuseness feature as an additional input to a DNN-based acoustic model leads to a reduced word error rate for the REVERB challenge corpus, both compared to logmelspec features extracted from noisy signals, and features enhanced by spectral subtraction.
Recently, phonetic posteriorgrams (PPGs) based methods have been quite popular in non-parallel singing voice conversion systems. However, due to the lack of acoustic information in PPGs, style and naturalness of the converted singing voices are still limited. To solve these problems, in this paper, we utilize an acoustic reference encoder to implicitly model singing characteristics. We experiment with different auxiliary features, including mel spectrograms, HuBERT, and the middle hidden feature (PPG-Mid) of pretrained automatic speech recognition (ASR) model, as the input of the reference encoder, and finally find the HuBERT feature is the best choice. In addition, we use contrastive predictive coding (CPC) module to further smooth the voices by predicting future observations in latent space. Experiments show that, compared with the baseline models, our proposed model can significantly improve the naturalness of converted singing voices and the similarity with the target singer. Moreover, our proposed model can also make the speakers with just speech data sing.
We present a state-of-the-art speech recognition system developed using end-to-end deep learning. Our architecture is significantly simpler than traditional speech systems, which rely on laboriously engineered processing pipelines; these traditional systems also tend to perform poorly when used in noisy environments. In contrast, our system does not need hand-designed components to model background noise, reverberation, or speaker variation, but instead directly learns a function that is robust to such effects. We do not need a phoneme dictionary, nor even the concept of a "phoneme." Key to our approach is a well-optimized RNN training system that uses multiple GPUs, as well as a set of novel data synthesis techniques that allow us to efficiently obtain a large amount of varied data for training. Our system, called Deep Speech, outperforms previously published results on the widely studied Switchboard Hub5'00, achieving 16.0% error on the full test set. Deep Speech also handles challenging noisy environments better than widely used, state-of-the-art commercial speech systems.
Affective computing is very important in the relationship between man and machine. In this paper, a system for speech emotion recognition (SER) based on speech signal is proposed, which uses new techniques in different stages of processing. The system consists of three stages: feature extraction, feature selection, and finally feature classification. In the first stage, a complex set of long-term statistics features is extracted from both the speech signal and the glottal-waveform signal using a combination of new and diverse features such as prosodic, spectral, and spectro-temporal features. One of the challenges of the SER systems is to distinguish correlated emotions. These features are good discriminators for speech emotions and increase the SER's ability to recognize similar and different emotions. This feature vector with a large number of dimensions naturally has redundancy. In the second stage, using classical feature selection techniques as well as a new quantum-inspired technique to reduce the feature vector dimensionality, the number of feature vector dimensions is reduced. In the third stage, the optimized feature vector is classified by a weighted deep sparse extreme learning machine (ELM) classifier. The classifier performs classification in three steps: sparse random feature learning, orthogonal random projection using the singular value decomposition (SVD) technique, and discriminative classification in the last step using the generalized Tikhonov regularization technique. Also, many existing emotional datasets suffer from the problem of data imbalanced distribution, which in turn increases the classification error and decreases system performance. In this paper, a new weighting method has also been proposed to deal with class imbalance, which is more efficient than existing weighting methods. The proposed method is evaluated on three standard emotional databases.
Performing joint interaction requires constant mutual monitoring of own actions and their effects on the other's behaviour. Such an action-effect monitoring is boosted by social cues and might result in an increasing sense of agency. Joint actions and joint attention are strictly correlated and both of them contribute to the formation of a precise temporal coordination. In human-robot interaction, the robot's ability to establish joint attention with a human partner and exploit various social cues to react accordingly is a crucial step in creating communicative robots. Along the social component, an effective human-robot interaction can be seen as a new method to improve and make the robot's learning process more natural and robust for a given task. In this work we use different social skills, such as mutual gaze, gaze following, speech and human face recognition, to develop an effective teacher-learner scenario tailored to visual object learning in dynamic environments. Experiments on the iCub robot demonstrate that the system allows the robot to learn new objects through a natural interaction with a human teacher in presence of distractors.
Distributed learning paradigms such as federated learning often involve transmission of model updates, or gradients, over a network, thereby avoiding transmission of private data. However, it is possible for sensitive information about the training data to be revealed from such gradients. Prior works have demonstrated that labels can be revealed analytically from the last layer of certain models (e.g., ResNet), or they can be reconstructed jointly with model inputs by using Gradients Matching [Zhu et al'19] with additional knowledge about the current state of the model. In this work, we propose a method to discover the set of labels of training samples from only the gradient of the last layer and the id to label mapping. Our method is applicable to a wide variety of model architectures across multiple domains. We demonstrate the effectiveness of our method for model training in two domains - image classification, and automatic speech recognition. Furthermore, we show that existing reconstruction techniques improve their efficacy when used in conjunction with our method. Conversely, we demonstrate that gradient quantization and sparsification can significantly reduce the success of the attack.
Deep learning (DL) has proven its unprecedented success in diverse fields such as computer vision, natural language processing, and speech recognition by its strong representation ability and ease of computation. As we move forward to a thoroughly intelligent society with 6G wireless networks, new applications and use-cases have been emerging with stringent requirements for next-generation wireless communications. Therefore, recent studies have focused on the potential of DL approaches in satisfying these rigorous needs and overcoming the deficiencies of existing model-based techniques. The main objective of this article is to unveil the state-of-the-art advancements in the field of DL-based physical layer (PHY) methods to pave the way for fascinating applications of 6G. In particular, we have focused our attention on four promising PHY concepts foreseen to dominate next-generation communications, namely massive multiple-input multiple-output (MIMO) systems, sophisticated multi-carrier (MC) waveform designs, reconfigurable intelligent surface (RIS)-empowered communications, and PHY security. We examine up-to-date developments in DL-based techniques, provide comparisons with state-of-the-art methods, and introduce a comprehensive guide for future directions. We also present an overview of the underlying concepts of DL, along with the theoretical background of well-known DL techniques. Furthermore, this article provides programming examples for a number of DL techniques and the implementation of a DL-based MIMO by sharing user-friendly code snippets, which might be useful for interested readers.
Automatic emotion recognition for real-life appli-cations is a challenging task. Human emotion expressions aresubtle, and can be conveyed by a combination of several emo-tions. In most existing emotion recognition studies, each audioutterance/video clip is labelled/classified in its entirety. However,utterance/clip-level labelling and classification can be too coarseto capture the subtle intra-utterance/clip temporal dynamics. Forexample, an utterance/video clip usually contains only a fewemotion-salient regions and many emotionless regions. In thisstudy, we propose to use attention mechanism in deep recurrentneural networks to detection the Regions-of-Interest (ROI) thatare more emotionally salient in human emotional speech/video,and further estimate the temporal emotion dynamics by aggre-gating those emotionally salient regions-of-interest. We comparethe ROI from audio and video and analyse them. We comparethe performance of the proposed attention networks with thestate-of-the-art LSTM models on multi-class classification task ofrecognizing six basic human emotions, and the proposed attentionmodels exhibit significantly better performance. Furthermore, theattention weight distribution can be used to interpret how anutterance can be expressed as a mixture of possible emotions.