Multimodal emotion recognition (MER) refers to the identification and understanding of human emotional states by combining different signals, including—but not limited to—text, speech, and face cues. MER plays a crucial role in the human–computer interaction (HCI) domain. With the recent progression of deep learning technologies and the increasing availability of multimodal datasets, the MER domain has witnessed considerable development, resulting in numerous significant research breakthroughs. However, a conspicuous absence of thorough and focused reviews on these deep learning-based MER achievements is observed. This survey aims to bridge this gap by providing a comprehensive overview of the recent advancements in MER based on deep learning. For an orderly exposition, this paper first outlines a meticulous analysis of the current multimodal datasets, emphasizing their advantages and constraints. Subsequently, we thoroughly scrutinize diverse methods for multimodal emotional feature extraction, highlighting the merits and demerits of each method. Moreover, we perform an exhaustive analysis of various MER algorithms, with particular focus on the model-agnostic fusion methods (including early fusion, late fusion, and hybrid fusion) and fusion based on intermediate layers of deep models (encompassing simple concatenation fusion, utterance-level interaction fusion, and fine-grained interaction fusion). We assess the strengths and weaknesses of these fusion strategies, providing guidance to researchers to help them select the most suitable techniques for their studies. In summary, this survey aims to provide a thorough and insightful review of the field of deep learning-based MER. It is intended as a valuable guide to aid researchers in furthering the evolution of this dynamic and impactful field.

Spectrogram is commonly used as the input feature of deep neural networks to learn the high(er)-level time–frequency pattern of speech signal for speech emotion recognition (SER). Generally, different emotions correspond to specific energy activations both within frequency bands and time frames on spectrogram, which indicates the frequency and time domains are both essential to represent the emotion for SER. However, recent spectrogram-based works mainly focus on modeling the long-term dependency in time domain, which makes these methods suffer from the following issues: 1) neglecting to model the emotion-related correlations within frequency domain during the time–frequency joint learning and 2) ignoring to capture the specific frequency bands associated with emotions. To cope with the issues, we propose an attentive time–frequency neural network (ATFNN) for SER, including a time–frequency neural network (TFNN) and time–frequency attention. Specifically, aiming at the first issue, we design a TFNN with a frequency-domain encoder (F-Encoder) based on the Transformer encoder and a time-domain encoder (T-Encoder) based on the bidirectional long short-term memory (Bi-LSTM). The F-Encoder and T-Encoder model the correlations within frequency bands and time frames, respectively, and they are embedded into a time–frequency joint learning strategy to obtain the time–frequency patterns of speech emotions. Moreover, to handle the second issue, we adopt the time–frequency attention with a frequency-attention network (F-Attention) and a time-attention network (T-Attention) to focus on the emotion-related long-range dependencies between frequency bands and across time frames, which can enhance the emotional discrimination of speech features. Extensive experimental results on three public emotional databases, i.e., IEMOCAP, ABC, and CASIA, show that our proposed ATFNN outperforms the state-of-the-art methods.

In this article, we propose a new unsupervised domain adaptation (DA) method called layer-adapted implicit distribution alignment networks (LIDANs) to address the challenge of cross-corpus speech emotion recognition (SER). LIDAN extends our previous ICASSP work, deep implicit distribution alignment networks (DIDANs), whose key contribution lies in the introduction of a novel regularization term called implicit distribution alignment (IDA). This term allows DIDAN trained on source (training) speech samples to remain applicable to predicting emotion labels for target (testing) speech samples, regardless of corpus variance in cross-corpus SER. To further enhance this method, we extend IDA to layer-adapted IDA (LIDA), resulting in LIDAN. This layer-adapted extension consists of three modified IDA terms that consider emotion labels at different levels of granularity. These terms are strategically arranged within different fully connected layers in LIDAN, aligning with the increasing emotion-discriminative abilities with respect to the layer depth. This arrangement enables LIDAN to more effectively learn emotion-discriminative and corpus-invariant features for SER across various corpora compared to DIDAN. It is also worthy to mention that unlike most existing methods that rely on estimating statistical moments to describe preassumed explicit distributions, both IDA and LIDA take a different approach. They utilize an idea of target sample reconstruction to directly bridge the feature distribution gap without making assumptions about their distribution type. As a result, DIDAN and LIDAN can be viewed as implicit cross-corpus SER methods. To evaluate LIDAN, we conducted extensive cross-corpus SER experiments on EmoDB, eNTERFACE, and CASIA corpora. The experimental results demonstrate that LIDAN surpasses recent state-of-theart explicit unsupervised DA methods in tackling cross-corpus SER tasks.

In this paper, we investigate the problem of cross-corpus speech emotion recognition (SER), in which the training (source) and testing (target) speech samples belong to different corpora. This case thus leads to a feature distribution mismatch between the source and target speech samples. Hence, the performance of most existing SER methods drops sharply. To solve this problem, we propose a simple yet effective transfer subspace learning method called joint distribution implicitly aligned subspace learning (JIASL). The basic idea of JIASL is very straightforward, i.e., building an emotion discriminative and corpus invariant linear regression model under an implicit distribution alignment strategy. Following this idea, we first make use of the source speech features and emotion labels to endow such a regression model with emotion-discriminative ability. Then, a well-designed reconstruction regularization term, jointly considering the marginal and conditional distribution alignments between the speech samples in both corpora, is adopted to implicitly enable the regression model to predict the emotion labels of target speech samples. To evaluate the performance of our proposed JIASL, extensive cross-corpus SER experiments are carried out, and the results demonstrate the promising performance of the proposed JIASL in coping with the tasks of cross-corpus SER.