Search for: All records

Building speech emotion recognition (SER) models for low-resource languages is challenging due to the scarcity of labeled speech data. This limitation mandates the development of cross-lingual unsupervised domain adaptation techniques to effectively utilize labeled data from resource-rich languages. Inspired by the TransVQA framework, we propose a method that leverages a shared quantized feature space to enable knowledge transfer between labeled and unlabeled data across languages. The approach utilizes a quantized codebook to capture shared features, while reducing the domain gap, and aligning class distributions, thereby improving classification accuracy. Additionally, an information loss (InfoLoss) mechanism mitigates critical information loss during quantization. InfoLoss achieves this goal by minimizing the loss within the simplex of posterior class label distributions. The proposed method demonstrates superior performance compared to state-of-the-art baseline approaches. Index Terms: Speech Emotion Recognition, Cross-lingual Unsupervised Domain Adaptation, Discrete Features, InfoLoss 

Unsupervised domain adaptation offers significant potential for cross-lingual speech emotion recognition (SER). Most relevant studies have addressed this problem as a domain mismatch without considering phonetical emotional differences across languages. Our study explores universal discrete speech units obtained with vector quantization of wavLM representations from emotional speech in English, Taiwanese Mandarin, and Russian. We estimate cluster-wise distributions of quantized wavLM frames to quantify phonetic commonalities and differences across languages, vowels, and emotions. Our findings indicate that certain emotion-specific phonemes exhibit cross-linguistic similarities. The distribution of vowels varies with emotional content. Certain vowels across languages show close distributional proximity, offering anchor points for cross-lingual domain adaptation. We also propose and validate a method to quantify phoneme distribution similarities across languages. 

The Interspeech 2025 speech emotion recognition in natural istic conditions challenge builds on previous efforts to advance speech emotion recognition (SER) in real-world scenarios. The focus is on recognizing emotions from spontaneous speech, moving beyond controlled datasets. It provides a framework for speaker-independent training, development, and evaluation, with annotations for both categorical and dimensional tasks. The challenge attracted 93 research teams, whose models significantly improved state-of-the-art results over competitive baselines. This paper summarizes the challenge, focusing on the key outcomes. We analyze top-performing methods, emerging trends, and innovative directions. We highlight the effectiveness of combining foundational models based on audio and text to achieve robust SER systems. The competition website, with leaderboards, baseline code, and instructions, is available at: https://lab-msp.com/MSP-Podcast_Competition/IS2025/. 

Cross-lingual speech emotion recognition (SER) is important for a wide range of everyday applications. While recent SER research relies heavily on large pretrained models for emotion training, existing studies often concentrate solely on the final transformer layer of these models. However, given the task-specific nature and hierarchical architecture of these models, each transformer layer encapsulates different levels of information. Leveraging this hierarchical structure, our study focuses on the information embedded across different layers. Through an examination of layer feature similarity across different languages, we propose a novel strategy called a layer-anchoring mechanism to facilitate emotion transfer in cross-lingual SER tasks. Our approach is evaluated using two distinct language affective corpora (MSP-Podcast and BIIC-Podcast), achieving a best UAR performance of 60.21% on the BIIC-podcast corpus. The analysis uncovers interesting insights into the behavior of popular pretrained models. 

The problem of predicting emotional attributes from speech has often focused on predicting a single value from a sentence or short speaking turn. These methods often ignore that natural emotions are both dynamic and dependent on context. To model the dynamic nature of emotions, we can treat the prediction of emotion from speech as a time-series problem. We refer to the problem of predicting these emotional traces as dynamic speech emotion recognition. Previous studies in this area have used models that treat all emotional traces as coming from the same underlying distribution. Since emotions are dependent on contextual information, these methods might obscure the context of an emotional interaction. This paper uses a neural process model with a segment-level speech emotion recognition (SER) model for this problem. This type of model leverages information from the time-series and predictions from the SER model to learn a prior that defines a distribution over emotional traces. Our proposed model performs 21% better than a bidirectional long short-term memory (BiLSTM) baseline when predicting emotional traces for valence. 

Emotion recognition is inherently a multimodal problem. Humans use both audible and visual cues to determine a person’s emotions. There has been extensive improvement in the methods we use to fuse audio and visual representations between two unimodal deep-learning models. However, there is a lack of accommodation for modalities that have a disparity in the amount of computational resources needed to provide the same amount of temporal information. As the sequence length increases, current methods often make simplifications such as discarding frames or cropping the sequence. This paper introduces a chunking methodology designed for cross-attention-based multimodal transformer architectures. The approach involves segmenting the visual input—the more computationally demanding modality—into chunks. Cross-attention is then performed between the encoded audio and visual features instead of the original sequence lengths of the unimodal backbones. Our method achieves significant improvements over conventional cross-attention techniques in the audio-visual domain for a six-class emotional recognition problem, demonstrating better F1 score, precision, and recall on the CREMA-D database while reducing computational overhead. 

Deep clustering is a popular unsupervised technique for feature representation learning. We recently proposed the chunk-based DeepEmoCluster framework for speech emotion recognition (SER) to adopt the concept of deep clustering as a novel semi-supervised learning (SSL) framework, which achieved improved recognition performances over conventional reconstruction-based approaches. However, the vanilla DeepEmoCluster lacks critical sentence- level temporal information that is useful for SER tasks. This study builds upon the DeepEmoCluster framework, creating a powerful SSL approach that leverages temporal information within a sentence. We propose two sentence-level temporal modeling alternatives using either the temporal-net or the triplet loss function, resulting in a novel temporal-enhanced DeepEmoCluster framework to capture essential temporal information. The key contribution to achieving this goal is the proposed sentence-level uniform sampling strategy, which preserves the original temporal order of the data for the clustering process. An extra network module (e.g., gated recurrent unit) is utilized for the temporal-net option to encode temporal information across the data chunks. Alternatively, we can impose additional temporal constraints by using the triplet loss function while training the DeepEmoCluster framework, which does not increase model complexity. Our experimental results based on the MSP-Podcast corpus demonstrate that the proposed temporal-enhanced framework significantly outperforms the vanilla DeepEmoCluster framework and other existing SSL approaches in regression tasks for the emotional attributes arousal, dominance, and valence. The improvements are observed in fully-supervised learning or SSL implementations. Further analyses validate the effectiveness of the proposed temporal modeling, showing (1) high temporal consistency in the cluster assignment, and (2) well-separated emotional patterns in the generated clusters. 

Most existing audio-text emotion recognition studies have focused on the computational modeling aspects, including strategies for fusing the modalities. An area that has received less attention is understanding the role of proper temporal synchronization between the modalities in the model performance. This study presents a transformer-based model designed with a word-chunk concept, which offers an ideal framework to explore different strategies to align text and speech. The approach creates chunks with alternative alignment strategies with different levels of dependency on the underlying lexical boundaries. A key contribution of this study is the multi-scale chunk alignment strategy, which generates random alignments to create the chunks without considering lexical boundaries. For every epoch, the approach generates a different alignment for each sentence, serving as an effective regularization method for temporal dependency. Our experimental results based on the MSP-Podcast corpus indicate that providing precise temporal alignment information to create the audio-text chunks does not improve the performance of the system. The attention mechanisms in the transformer-based approach are able to compensate for imperfect synchronization between the modalities. However, using exact lexical boundaries makes the system highly vulnerable to missing modalities. In contrast, the model trained with the proposed multi-scale chunk regularization strategy using random alignment can significantly increase its robustness against missing data and remain effective, even under a single audio-only emotion recognition task. The code is available at: https://github.com/winston-lin-wei-cheng/MultiScale-Chunk-Regularization