EmotionThinker: LLMs learn to explain emotions in speech

For a long time, emotion recognition systems worked as a black box: the model would listen to a voice, output a label — "sadness," "joy," "anger" — and that was the end of its work. No explanations, no context. A group of researchers decided to fundamentally change this approach, and the result of their work — the EmotionThinker model — was presented at the ICLR 2026 conference as an Oral presentation, which in itself testifies to the high assessment of the scientific community.

The essence of the problem that EmotionThinker solves is easy to understand. Traditional emotion recognition in speech is a classification task: the system learns to match acoustic features with a predefined set of emotional categories. The approach works, but has a fundamental flaw — lack of transparency. A clinical psychologist listening to a patient doesn't simply label them as "anxious." He notices voice trembling on certain words, pauses where you wouldn't expect them, acceleration of speech rate at specific moments. EmotionThinker for the first time transfers this analytical process into the space of large language models.

Architecturally, the model is built on the idea of a chain of thought — an approach that has become one of the key directions in LLM development over the past two years. Instead of immediately producing a classification answer, EmotionThinker first generates a detailed text explanation: why this particular emotion, what acoustic and semantic signals point to it, how the meaning of the words being spoken and the manner of their delivery interact with each other. Only after this step does the model formulate the final conclusion. It is fundamentally important that the explanation is not an ex post facto rationalization, but a direct part of the decision-making process.

This is where the main technical achievement lies. Speech signals and text are fundamentally different modalities, and their joint processing remains one of the most challenging tasks in multimodal AI. Speech carries information that cannot be conveyed in words: intonational rises, micro-pauses, timbral changes. EmotionThinker learns not simply to translate these signals into text, but to construct a coherent narrative connecting the acoustic layer with the semantic layer. This is what makes the model's explanations substantive rather than formal.

The significance of this work extends far beyond academic interest. Transparency in emotional AI is a matter of trust and applicability. Imagine a mental health support system that doesn't just detect anxiety markers in a user's voice, but can explain to an operator or the user themselves what exactly alerted the algorithm. Or a quality control system in a call center that doesn't just flag a call, but points to specific moments where the emotional tone of the interaction began to degrade. In education, in medicine, in corporate communications — everywhere that matters not only what a person feels but why, such systems acquire qualitatively different value.

Also important is the broader context. The European AI Act, adopted in 2024, introduces strict requirements for explainability of systems operating in sensitive areas. Emotion recognition is one of them. The EmotionThinker approach fits organically into this regulatory trend: a model that can explain its decisions fits much more easily into audit and verification requirements. The researchers have essentially proposed an architectural answer to a legal challenge.

Of course, open questions remain. To what extent do the generated explanations truly reflect the model's internal logic, rather than being plausible but arbitrary texts — an issue that will require independent research to verify. Moreover, the generalizability of the approach across different languages and cultural contexts — where norms of emotional expression differ fundamentally — will require separate work. The Chinese audience, for which the system was originally created, and, say, the Mediterranean — are completely different emotional environments.

Nevertheless, EmotionThinker marks an important direction. Emotion recognition ceases to be a classification task and becomes an understanding task. AI that can not only feel but also explain — this is a fundamentally different level of human-machine interaction. And the fact that this work received Oral status at ICLR 2026 indicates that the scientific community understands this.