Speech technologies have been developed for decades as a typical signal processing area, while the last decade has brought a huge progress based on new machine learning paradigms. Owing not only to their intrinsic complexity but also to their relation with cognitive sciences, speech technologies are now viewed as a prime example of interdisciplinary knowledge area. This review article on speech signal analysis and processing, corresponding machine learning algorithms, and applied computational intelligence aims to give an insight into several fields, covering speech production and auditory perception, cognitive aspects of speech communication and language understanding, both speech recognition and text-to-speech synthesis in more details, and consequently the main directions in development of spoken dialogue systems. Additionally, the article discusses the concepts and recent advances in speech signal compression, coding, and transmission, including cognitive speech coding. To conclude, the main intention of this article is to highlight recent achievements and challenges based on new machine learning paradigms that, over the last decade, had an immense impact in the field of speech signal processing.

Human–machine interaction covers a range of applications in which machines should understand humans’ commands and predict their behavior. Humans commonly change their mood over time, which affects the way we interact, particularly by changing speech style and facial expressions. As interaction requires quick decisions, low latency is critical for real-time processing. Edge devices, strategically placed near the data source, minimize processing time, enabling real-time decision-making. Edge computing allows us to process data locally, thus reducing the need to send sensitive information further through the network. Despite the wide adoption of audio-only, video-only, and multimodal emotion recognition systems, there is a research gap in terms of analyzing lightweight models and solving privacy challenges to improve model performance. This motivated us to develop a privacy-preserving, lightweight, CNN-based (CNNs are frequently used for processing audio and video modalities) audiovisual emotion recognition model, deployable on constrained edge devices. The model is further paired with a federated learning protocol to preserve the privacy of local clients on edge devices and improve detection accuracy. The results show that the adoption of federated learning improved classification accuracy by ~2%, as well as that the proposed federated learning-based model provides competitive performance compared to other baseline audiovisual emotion recognition models.

Speaker recognition is an important classification task, which can be solved using several approaches. Although building a speaker recognition model on a closed set of speakers under neutral speaking conditions is a well-researched task and there are solutions that provide excellent performance, the classification accuracy of developed models significantly decreases when applying them to emotional speech or in the presence of interference. Furthermore, deep models may require a large number of parameters, so constrained solutions are desirable in order to implement them on edge devices in the Internet of Things systems for real-time detection. The aim of this paper is to propose a simple and constrained convolutional neural network for speaker recognition tasks and to examine its robustness for recognition in emotional speech conditions. We examine three quantization methods for developing a constrained network: floating-point eight format, ternary scalar quantization, and binary scalar quantization. The results are demonstrated on the recently recorded SEAC dataset.

Gaussian Mixture Models (GMMs) are used in many traditional expert systems and modern artificial intelligence tasks such as automatic speech recognition, image recognition and retrieval, pattern recognition, speaker recognition and verification, financial forecasting applications and others, as simple statistical representations of underlying data. Those representations typically require many high-dimensional GMM components that consume large computing resources and increase computation time. On the other hand, real-time applications require computationally efficient algorithms and for that reason, various GMM similarity measures and dimensionality reduction techniques have been examined to reduce the computational complexity. In this paper, a novel GMM similarity measure is proposed. The measure is based on a recently presented nonlinear geometry-aware dimensionality reduction algorithm for the manifold of Symmetric Positive Definite (SPD) matrices. The algorithm is applied over SPD representations of the original data. The local neighborhood information from the original high-dimensional parameter space is preserved by preserving distance to the local mean. Instead of dealing with high-dimensional parameter space, the method operates on much lower-dimensional space of transformed parameters. Resolving the distance between such representations is reduced to calculating the distance among lower-dimensional matrices. The method was tested within a texture recognition task where superior state-of-the-art performance in terms of the trade-off between recognition accuracy and computational complexity has been achieved in comparison with all baseline GMM similarity measures.

Achieving real-time inference is one of the major issues in contemporary neural network applications, as complex algorithms are frequently being deployed to mobile devices that have constrained storage and computing power. Moving from a full-precision neural network model to a lower representation by applying quantization techniques is a popular approach to facilitate this issue. Here, we analyze in detail and design a 2-bit uniform quantization model for Laplacian source due to its significance in terms of implementation simplicity, which further leads to a shorter processing time and faster inference. The results show that it is possible to achieve high classification accuracy (more than 96% in the case of MLP and more than 98% in the case of CNN) by implementing the proposed model, which is competitive to the performance of the other quantization solutions with almost optimal precision.

This paper considers the design of a binary scalar quantizer of Laplacian source and its application in compressed neural networks. The quantizer performance is investigated in a wide dynamic range of data variances, and for that purpose, we derive novel closed-form expressions. Moreover, we propose two selection criteria for the variance range of interest. Binary quantizers are further implemented for compressing neural network weights and its performance is analysed for a simple classification task. Good matching between theory and experiment is observed and a great possibility for implementation is indicated.