Multimodal Emotion Recognition (MER) systems often suffer from missing modalities in real-world scenarios. Existing methods usually generate, align, or distill missing modalities as a whole, overlooking the heterogeneous nature of the information carried by each modality. Such holistic treatment mixes inferable shared semantics with uncertain modality-specific details, yielding unstable representations and degrading robustness. To address this issue, we propose the Primitive Memory Distillation (PriMD) framework. Unlike existing methods, PriMD takes an intra-modal perspective and focuses on how different types of information within a modality differ in recoverability within each modality. PriMD first disentangles cross-modal shared semantics from modality-specific representations, and then discretizes the latter into learnable semantic primitives to construct modality-specific memory banks. When modalities are missing, PriMD is a teacher–student framework that the student model uses the shared semantics of available modalities as queries to dynamically retrieve primitives. It compensates for missing modality-specific information within a constrained memory space and aligns with the teacher model. Extensive experiments on IEMOCAP, CMU-MOSI, and CMU-MOSEI demonstrate that PriMD achieves state-of-the-art performance and consistently stronger robustness across a wide range of missing-modality settings, while mitigating the instability caused by holistic feature inference. Our code and project website are available at https://github.com/JiaqiZhang-Sengoku/PriMD and https://jiaqizhang-sengoku.github.io/PriMD/, respectively.

Speech-generative large audio-language models (LALMs) can produce expressive speech, yet reliable emotion control remains elusive: conversions often miss the target affect and may degrade linguistic fidelity through refusals, hallucinations, or paraphrase. We present, to our knowledge, one of the first neuron-level studies of emotional voice conversion (EVC) in LALMs and provide intervention-based evidence that compact sets of emotion-sensitive neuron (ESN) are functionally involved in, and can be exploited for, steering expressive speech generation at inference time. ESNs are identified via success-filtered activation aggregation enforcing both emotion realization and content preservation. Across three LALMs (Qwen2.5-Omni-7B, MiniCPM-o 4.5, Kimi-Audio), ESN interventions yield emotion-specific gains that generalize to unseen speakers. Controllability depends on selector design, mask sparsity, filtering, and intervention strength. Our results provide an intervention-based framework for training-free emotion control in LALMs for EVC.

Large Audio Language Models (LALMs) have shown strong performance on audio reasoning benchmarks, but accuracy alone cannot distinguish true reasoning from superficial pattern matching, often overestimating reasoning ability since high scores may result from guessing rather than genuine audio understanding. Evaluating the reasoning process itself is essential for improving LALMs’ reasoning ability, yet remains challenging. Existing methods either rely on costly human annotation or opaque LLM-as-judge approaches, making them impractical, biased, and lacking transparency. Moreover, audio reasoning introduces unique challenges absent in text-based settings, perceptual hallucination and cross-modal alignment between audio understanding and textual inference, hence text-based evaluation frameworks cannot be directly applied. Therefore, we propose ARIA-Rubrics (Audio Reasoning Integrity Assessment), a lightweight, annotation-free gold reasoning chains, automatic and transparent framework comprising six complementary metrics that evaluate audio reasoning quality across perceptual grounding, reasoning coherence, and answer consistency. We use Chain-of-Thought prompting as an externalization mechanism to make the reasoning process observable. Experiments on 9 models across 2 benchmarks identify three reasoning modes of current LALMs with actionable directions for future development, with ARIA-Rubrics achieving high correlation with human judgments. The code is available at the https://github.com/glam-imperial/allm_assessment.

Previous work links Major Depressive Disorder (MDD) to linguistic and paralinguistic speech patterns, highlighting speech analysis as a potential diagnostic tool for remote monitoring or clinical decision support (CDS) systems. Recent attempts to build MDD diagnostic models primarily rely on the English (E-)DAIC corpus. Despite promising results, these works rarely evaluate cross-corpus generalization and rely on self-reported questionnaires, limiting their clinical utility. We introduce a new, clinically validated German corpus, EmpkinS-EKSpression, of semi-structured diagnostic interviews to test the reliability of spoken language-based measures of MDD. Our model achieves an F1-score of 0.86 on the EmpkinS-EKSpression dataset and 0.74 on E-DAIC via zero-shot cross-corpus transfer, suggesting that MDD detection models can generalize across corpora and languages despite differences in interview format and ground truth definition, supporting further development of CDS systems.

Emotion is a central dimension of spoken communication, yet, we still lack a mechanistic account of how modern large audio-language models (LALMs) encode it internally. We present the first neuron-level interpretability study of emotion-sensitive neurons (ESNs) in LALMs and provide causal evidence supporting the existence of such units in Qwen2.5-Omni, Kimi-Audio, and Audio Flamingo 3. Across these three widely used open-source models, we compare frequency-, entropy-, mean-deviation-, and contrast-based neuron selectors on multiple emotion recognition benchmarks. Using inference-time interventions, we reveal a consistent emotion-specific signature: deactivating neurons selected for a given emotion disproportionately degrades recognition of that emotion while largely preserving other classes, whereas targeted steering amplifies these units to bias predictions toward the target emotion. These effects arise with modest amounts of identification data and scale systematically with intervention strength. We further observe that ESNs exhibit non-uniform layer-wise clustering with partial cross-dataset transfer. Taken together, our results offer a causal, neuron-level account of emotion decisions in LALMs and highlight targeted neuron interventions as an actionable handle for controllable affective behaviors.

Robust detection of AI-generated text under deployment conditions is challenging: distribution shifts across domains and generators, adversarial perturbations of the input surface, and the absence of target-domain labels for threshold calibration all degrade detectors that perform well in-domain. We present DeBERTa-ConPara, a deployment-oriented detector combining attack-aware Unicode preprocessing with a contextual transformer encoder trained over HC3 Plus, M4, MAGE and RAID. Our central finding is that preprocessing acts in opposite directions depending on where it is applied: normalising the training corpus deduplicates it, collapsing 35.4% of RAID rows into copies of their clean siblings and deleting the adversarial supervision, whereas normalising at inference is an effective defence. A factorial varying the two placements independently identifies raw training with normalised inference as the best configuration, reaching 99.61% AUROC, 99.01% TPR@5% FPR and 96.57% TPR@1% FPR on the official RAID hidden test, alongside 93.14% average balanced accuracy across HC3 Plus and MAGE under a fixed threshold. The gain is confined to two of twelve attack classes: homoglyph and zero-width-space insertion rise from 11.05% and 1.12% to 96.98%. The same signature reproduces in a zero-shot detector of different architecture, showing the effect belongs to the attacks rather than to our model. We additionally report two negative results: semantic-invariance augmentation through paraphrasing and supervised contrastive learning (ConPara) does not improve the best configuration, and the handcrafted feature-fusion branch is inert in distribution and harmful outside it.

Speaking-style captioning (SSC) aims to generate natural language descriptions of how speech is delivered, capturing paralinguistic attributes such as vocal timbre, prosody, and expressivity. Many downstream applications, including conversational AI agents, controllable speech generation, and large-scale audio annotation, require dimension-specific style captions that describe targeted aspects of speech (e.g., speaker traits, emotion, or delivery style), rather than a single undifferentiated description. However, existing work lacks a unified formulation for controllable speaking-style captioning. We address this gap by formulating SSC as an instruction-following audio-language task, where explicit instructions specify the speaking-style dimensions to be described. We introduce StyleInstructCaps, the first standardized benchmark for controllable speaking-style captioning in audio-language models. StyleInstructCaps provides a dataset and evaluation framework that measures metadata groundedness, hallucination, instruction-following ability, generalization to unseen instructions and audio datasets, and speaker-style consistency. Across nine state-of-the-art audio-language models, we find that strong instruction-following ability does not necessarily imply grounded speaking-style understanding, and that SSC-specific instruction-diverse training substantially improves groundedness, controllability, and speaker consistency.

Telling stories is an integral part of human communication which can evoke emotions and influence the affective states of the audience. Automatically modeling emotional trajectories in stories has thus attracted considerable scholarly interest. However, as most existing works have been limited to unsupervised dictionary-based approaches, there is no benchmark for this task. We address this gap by introducing continuous valence and arousal labels for an existing dataset of children’s stories originally annotated with discrete emotion categories. We collect additional annotations for this data and map the categorical labels to the continuous valence and arousal space. For predicting the thus obtained emotionality signals, we fine-tune a DeBERTa model and improve upon this baseline via a weakly supervised learning approach. The best configuration achieves a Concordance Correlation Coefficient (CCC) of .8221 for valence and .7125 for arousal on the test set, demonstrating the efficacy of our proposed approach. A detailed analysis shows the extent to which the results vary depending on factors such as the author, the individual story, or the section within the story. In addition, we uncover the weaknesses of our approach by investigating examples that prove to be difficult to predict.

We present advancements with a software tool called Nkululeko, that lets users perform (semi-) supervised machine learning experiments in the speaker characteristics domain. It is based on audformat, a format for speech database metadata description. Due to an interface based on configurable templates, it supports best practise and very fast setup of experiments without the need to be proficient in the underlying language: Python. The paper explains the handling of Nkululeko and presents two typical experiments: comparing the expert acoustic features with artificial neural net embeddings for emotion classification and speaker age regression.

Since several decades emotional databases have been recorded by various laboratories. Many of them contain acted portrays of Darwin’s famous “big four” basic emotions. In this paper, we investigate in how far a selection of them are comparable by two approaches: on the one hand modeling similarity as performance in cross database machine learning experiments and on the other by analyzing a manually picked set of four acoustic features that represent different phonetic areas. It is interesting to see in how far specific databases (we added a synthetic one) perform well as a training set for others while some do not. Generally speaking, we found indications for both similarity as well as specificiality across languages.

Speech emotion recognition is in the focus of research since several decades and has many applications. One problem is sparse data for supervised learning. One way to tackle this problem is the synthesis of data with emotion simulating speech synthesis approaches. We present a synthesized database of five basic emotions and neutral expression based on rule based manipulation for a diphone synthesizer which we release to the public. The database has been validated in several machine learning experiments as a training set to detect emotional expression from natural speech data. The scripts to generate such a database have been made open source and could be used to aid speech emotion recognition for a low resourced language, as MBROLA supports 35 languages

Traditional learning-based approaches to sentiment analysis of written text use the concept of bag-of-words or bag-of-n-grams, where a document is viewed as a set of terms or short combinations of terms disregarding grammar rules or word order. Novel approaches de-emphasize this concept and view the problem as a sequence classification problem. In this context, recurrent neural networks (RNNs) have achieved significant success. The idea is to use RNNs as discriminative binary classifiers to predict a positive or negative sentiment label at every word position then perform a type of pooling to get a sentence-level polarity. Here, we investigate a novel generative approach in which a separate probability distribution is estimated for every sentiment using language models (LMs) based on long short-term memory (LSTM) RNNs. We introduce a novel type of LM using a modified version of bidirectional LSTM (BLSTM) called contextual BLSTM (cBLSTM), where the probability of a word is estimated based on its full left and right contexts. Our approach is compared with a BLSTM binary classifier. Significant improvements are observed in classifying the IMDB movie review dataset. Further improvements are achieved via model combination.

Crowdsourcing is an arising collaborative approach applicable among many other applications to the area of language and speech processing. In fact, the use of crowdsourcing was already applied in the field of speech processing with promising results. However, only few studies investigated the use of crowdsourcing in computational paralinguistics. In this contribution, we propose a novel evaluator for crowdsourced-based ratings termed Weighted Trustability Evaluator (WTE) which is computed from the rater-dependent consistency over the test questions. We further investigate the reliability of crowdsourced annotations as compared to the ones obtained with traditional labelling procedures, such as constrained listening experiments in laboratories or in controlled environments. This comparison includes an in-depth analysis of obtainable classification performances. The experiments were conducted on the Speaker Likability Database (SLD) already used in the INTERSPEECH Challenge 2012, and the results lend further weight to the assumption that crowdsourcing can be applied as a reliable annotation source for computational paralinguistics given a sufficient number of raters and suited measurements of their reliability.

In this paper, we describe a new database with audio recordings of non-native (L2) speakers of English, and the perceptual evaluation experiment conducted with native English speakers for assessing the prosody of each recording. These annotations are then used to compute the gold standard using different methods, and a series of regression experiments is conducted to evaluate their impact on the performance of a regression model predicting the degree of naturalness of L2 speech. Further, we compare the relevance of different feature groups modelling prosody in general (without speech tempo), speech rate and pauses modelling speech tempo (fluency), voice quality, and a variety of spectral features. We also discuss the impact of various fusion strategies on performance. Overall, our results demonstrate that the prosody of non-native speakers of English as L2 can be reliably assessed using supra-segmental audio features; prosodic features seem to be the most important ones.

An important difference between traditional AI systems and human intelligence is the human ability to harness commonsense knowledge gleaned from a lifetime of learning and experience to make informed decisions. This allows humans to adapt easily to novel situations where AI fails catastrophically due to a lack of situation-specific rules and generalization capabilities. Commonsense knowledge also provides background information that enables humans to successfully operate in social situations where such knowledge is typically assumed. Since commonsense consists of information that humans take for granted, gathering it is an extremely difficult task. Previous versions of SenticNet were focused on collecting this kind of knowledge for sentiment analysis but they were heavily limited by their inability to generalize. SenticNet 4 overcomes such limitations by leveraging on conceptual primitives automatically generated by means of hierarchical clustering and dimensionality reduction.

We introduce a spoken language resource for the analysis of impact that physical exercising has on human speech production. In particular, the database provides heart rate and skin conductance measurement information alongside the audio recordings. It contains recordings from 19 subjects in a relaxed state and after exercising. The audio material includes breathing, sustained vowels, and read text. Further, we describe pre-extracted audio-features from our openSMILE feature extractor together with baseline performances for the recognition of high and low heart rate using these features. The baseline results clearly show the feasibility of automatic estimation of heart rate from the human voice, in particular from sustained vowels. Both regression - in order to predict the exact heart rate value - and a binary classification setting for high and low heart rate classes are investigated. Finally, we give tendencies on feature group relevance in the named contexts of heart rate estimation and skin conductivity estimation.

The CINEMO corpus of French emotional speech provides a richly annotated resource to help overcome the apparent lack of learning and testing speech material for complex, i.e. blended or mixed emotions. The protocol for its collection was dubbing selected emotional scenes from French movies. 51 speakers are contained and the total speech time amounts to 2 hours and 13 minutes and 4k speech chunks after segmentation. Extensive labelling was carried out in 16 categories for major and minor emotions and in 6 continuous dimensions. In this contribution we give insight into the corpus statistics focusing in particular on the topic of complex emotions, and provide benchmark recognition results obtained in exemplary large feature space evaluations. In the result the labelling oft he collected speech clearly demonstrates that a complex handling of emotion seems needed. Further, the automatic recognition experiments provide evidence that the automatic recognition of blended emotions appears to be feasible.