Emotional speech synthesis aims to synthesize human voices with various emotional effects. The current studies are mostly focused on imitating an averaged style belonging to a specific emotion type. In this paper, we seek to generate speech with a mixture of emotions at run-time. We propose a novel formulation that measures the relative difference between the speech samples of different emotions. We then incorporate our formulation into a sequence-to-sequence emotional text-to-speech framework. During the training, the framework does not only explicitly characterize emotion styles, but also explores the ordinal nature of emotions by quantifying the differences with other emotions. At run-time, we control the model to produce the desired emotion mixture by manually defining an emotion attribute vector. The objective and subjective evaluations have validated the effectiveness of the proposed framework. To our best knowledge, this research is the first study on modelling, synthesizing and evaluating mixed emotions in speech.
FullSubNet is our recently proposed real-time single-channel speech enhancement network that achieves outstanding performance on the Deep Noise Suppression (DNS) Challenge dataset. A number of variants of FullSubNet have been proposed recently, but they all focus on the structure design towards better performance and are rarely concerned with computational efficiency. This work proposes a new architecture named Fast FullSubNet dedicated to accelerating the computation of FullSubNet. Specifically, Fast FullSubNet processes sub-band speech spectra in the mel-frequency domain by using cascaded linear-to-mel full-band, sub-band, and mel-to-linear full-band models such that frequencies involved in the sub-band computation are vastly reduced. After that, a down-sampling operation is proposed for the sub-band input sequence to further reduce the computational complexity along the time axis. Experimental results show that, compared to FullSubNet, Fast FullSubNet has only 13% computational complexity and 16% processing time, and achieves comparable or even better performance.
The performance of data-driven natural language processing systems is contingent upon the quality of corpora. However, principal corpus design criteria are often not identified and examined adequately, particularly in the speech processing discipline. Speech corpora development requires additional attention with regard to clean/noisy, read/spontaneous, multi-talker speech, accents/dialects, etc. Domain selection is also a crucial decision point in speech corpus development. In this study, we demonstrate the significance of domain selection by assessing a state-of-the-art Bangla automatic speech recognition (ASR) model on a novel multi-domain Bangladeshi Bangla ASR evaluation benchmark - BanSpeech, which contains 7.2 hours of speech and 9802 utterances from 19 distinct domains. The ASR model has been trained with deep convolutional neural network (CNN), layer normalization technique, and Connectionist Temporal Classification (CTC) loss criterion on SUBAK.KO, a mostly read speech corpus for the low-resource and morphologically rich language Bangla. Experimental evaluation reveals the ASR model on SUBAK.KO faces difficulty recognizing speech from domains with mostly spontaneous speech and has a high number of out-of-vocabulary (OOV) words. The same ASR model, on the other hand, performs better in read speech domains and contains fewer OOV words. In addition, we report the outcomes of our experiments with layer normalization, input feature extraction, number of convolutional layers, etc., and set a baseline on SUBAK.KO. The BanSpeech will be publicly available to meet the need for a challenging evaluation benchmark for Bangla ASR.
In this work, we present Slimmable Neural Networks applied to the problem of small-footprint keyword spotting. We show that slimmable neural networks allow us to create super-nets from Convolutioanl Neural Networks and Transformers, from which sub-networks of different sizes can be extracted. We demonstrate the usefulness of these models on in-house Alexa data and Google Speech Commands, and focus our efforts on models for the on-device use case, limiting ourselves to less than 250k parameters. We show that slimmable models can match (and in some cases, outperform) models trained from scratch. Slimmable neural networks are therefore a class of models particularly useful when the same functionality is to be replicated at different memory and compute budgets, with different accuracy requirements.
Compared to conventional artificial neurons that produce dense and real-valued responses, biologically-inspired spiking neurons transmit sparse and binary information, which can also lead to energy-efficient implementations. Recent research has shown that spiking neural networks can be trained like standard recurrent neural networks using the surrogate gradient method. They have shown promising results on speech command recognition tasks. Using the same technique, we show that they are scalable to large vocabulary continuous speech recognition, where they are capable of replacing LSTMs in the encoder with only minor loss of performance. This suggests that they may be applicable to more involved sequence-to-sequence tasks. Moreover, in contrast to their recurrent non-spiking counterparts, they show robustness to exploding gradient problems without the need to use gates.
The use of modern vocoders in an analysis/synthesis pipeline allows us to investigate high-quality voice conversion that can be used for privacy purposes. Here, we propose to transform the speaker embedding and the pitch in order to hide the sex of the speaker. ECAPA-TDNN-based speaker representation fed into a HiFiGAN vocoder is protected using a neural-discriminant analysis approach, which is consistent with the zero-evidence concept of privacy. This approach significantly reduces the information in speech related to the speaker's sex while preserving speech content and some consistency in the resulting protected voices.
The scarcity of labeled far-field speech is a constraint for training superior far-field speaker verification systems. Fine-tuning the model pre-trained on large-scale near-field speech substantially outperforms training from scratch. However, the fine-tuning method suffers from two limitations--catastrophic forgetting and overfitting. In this paper, we propose a weight transfer regularization(WTR) loss to constrain the distance of the weights between the pre-trained model with large-scale near-field speech and the fine-tuned model through a small number of far-field speech. With the WTR loss, the fine-tuning process takes advantage of the previously acquired discriminative ability from the large-scale near-field speech without catastrophic forgetting. Meanwhile, we use the PAC-Bayes generalization theory to analyze the generalization bound of the fine-tuned model with the WTR loss. The analysis result indicates that the WTR term makes the fine-tuned model have a tighter generalization upper bound. Moreover, we explore three kinds of norm distance for weight transfer, which are L1-norm distance, L2-norm distance and Max-norm distance. Finally, we evaluate the effectiveness of the WTR loss on VoxCeleb (pre-trained dataset) and FFSVC (fine-tuned dataset) datasets.
The rise of hate speech on online platforms has led to an urgent need for effective content moderation. However, the subjective and multi-faceted nature of hateful online content, including implicit hate speech, poses significant challenges to human moderators and content moderation systems. To address this issue, we developed ToxVis, a visually interactive and explainable tool for classifying hate speech into three categories: implicit, explicit, and non-hateful. We fine-tuned two transformer-based models using RoBERTa, XLNET, and GPT-3 and used deep learning interpretation techniques to provide explanations for the classification results. ToxVis enables users to input potentially hateful text and receive a classification result along with a visual explanation of which words contributed most to the decision. By making the classification process explainable, ToxVis provides a valuable tool for understanding the nuances of hateful content and supporting more effective content moderation. Our research contributes to the growing body of work aimed at mitigating the harms caused by online hate speech and demonstrates the potential for combining state-of-the-art natural language processing models with interpretable deep learning techniques to address this critical issue. Finally, ToxVis can serve as a resource for content moderators, social media platforms, and researchers working to combat the spread of hate speech online.
The careful construction of audio representations has become a dominant feature in the design of approaches to many speech tasks. Increasingly, such approaches have emphasized "disentanglement", where a representation contains only parts of the speech signal relevant to transcription while discarding irrelevant information. In this paper, we construct a representation learning task based on joint modeling of ASR and TTS, and seek to learn a representation of audio that disentangles that part of the speech signal that is relevant to transcription from that part which is not. We present empirical evidence that successfully finding such a representation is tied to the randomness inherent in training. We then make the observation that these desired, disentangled solutions to the optimization problem possess unique statistical properties. Finally, we show that enforcing these properties during training improves WER by 24.5% relative on average for our joint modeling task. These observations motivate a novel approach to learning effective audio representations.
Social media is a modern person's digital voice to project and engage with new ideas and mobilise communities $\unicode{x2013}$ a power shared with extremists. Given the societal risks of unvetted content-moderating algorithms for Extremism, Radicalisation, and Hate speech (ERH) detection, responsible software engineering must understand the who, what, when, where, and why such models are necessary to protect user safety and free expression. Hence, we propose and examine the unique research field of ERH context mining to unify disjoint studies. Specifically, we evaluate the start-to-finish design process from socio-technical definition-building and dataset collection strategies to technical algorithm design and performance. Our 2015-2021 51-study Systematic Literature Review (SLR) provides the first cross-examination of textual, network, and visual approaches to detecting extremist affiliation, hateful content, and radicalisation towards groups and movements. We identify consensus-driven ERH definitions and propose solutions to existing ideological and geographic biases, particularly due to the lack of research in Oceania/Australasia. Our hybridised investigation on Natural Language Processing, Community Detection, and visual-text models demonstrates the dominating performance of textual transformer-based algorithms. We conclude with vital recommendations for ERH context mining researchers and propose an uptake roadmap with guidelines for researchers, industries, and governments to enable a safer cyberspace.