Can AI Detect When Two People Are Emotionally in Sync? The TRACE Framework and What It Means for Hiring and Customer Support

If you’ve spent time in hiring, you’ve probably noticed this.

Some interviewers have a gift for it. Candidates who walk in tense and guarded gradually open up. Their answers get longer, more honest, more considered. The conversation starts to feel like a real exchange. Other interviewers, equally experienced, seem to leave candidates just as closed-off at the end as they were at the start.

The same thing plays out in call centers. Veteran operators somehow manage to turn frustrated customers around over the course of a single call. New operators, by contrast, sometimes end calls with both parties feeling worse than when they started. The customer’s frustration bled into the operator’s voice, the operator’s discomfort amplified the customer’s, and the whole thing spiraled.

We’ve all seen this pattern, and we’ve largely accepted that the skill behind it — the ability to emotionally sync with another person in real time — is something intuitive, difficult to measure, and harder to teach.

That assumption may be changing.

---

Three things this article covers

• A research framework called TRACE achieves 97.01% accuracy in detecting emotional entrainment — the phenomenon of two speakers’ emotional states aligning over the course of a conversation — from audio alone
• The key finding is that temporal context and speaker relationship information are both essential: stripping either degrades performance, confirming that emotional sync is fundamentally a time-evolving, relational process
• The business cases for automatic entrainment scoring — interviewer evaluation, call center quality monitoring, counseling support AI — are now within realistic technical reach

---

① What emotional entrainment actually is

Start with the phenomenon itself, because it’s easy to misread.

Emotional entrainment isn’t just about matching someone’s energy or mirroring their words. It refers to the process by which two people’s emotional states gradually converge during a conversation. One person’s emotional trajectory begins to influence the other’s, and over time the two move together — or don’t.

This is different from a single moment of empathy or rapport. It unfolds across the arc of a conversation. It requires tracking not just how each speaker is feeling at any given point, but how those states are changing in relation to each other over time.

That’s what makes it hard to detect computationally. A snapshot of vocal features at one moment won’t capture it. You need to model the temporal flow, and you need to track how the two speakers are interacting with each other — not just what each is doing independently.

---

② How TRACE approaches the problem

The research introduces two things: a dataset and a framework.

The dataset, called DyadEE, addresses a basic challenge in building this kind of system. To train and evaluate a detector, you need examples of both entrained and non-entrained conversations. Real-world conversational data contains plenty of natural emotional alignment, but almost no examples where that alignment has been deliberately disrupted. So the researchers built DyadEE to include synthetically altered conversations where emotional synchrony is artificially broken — creating a dataset where both sides of the detection problem are represented.

The framework itself — TRACE — extracts acoustic embeddings from speech representations generated by an emotion-tuned version of Whisper, the speech recognition model. These embeddings are arranged as ordered sequences, and the model then incorporates two additional layers of information: conversational context over time, and information about the relationship between the two speakers.

This combination — temporal order plus relational structure — is the core design choice. And it’s the one the experiments show matters most.

---

③ What 97.01% accuracy tells us

The TRACE framework achieves 97.01% accuracy in detecting emotional entrainment on the DyadEE dataset.

The number is striking, but what’s more informative is what the ablation results show. When the researchers removed temporal context from the model, accuracy dropped. When they removed speaker relationship information, accuracy dropped again. Both components, independently, are necessary for the system to work well.

This has a clear conceptual interpretation. Emotional entrainment is not just a property of how someone sounds at a given moment. It is a property of how two people are moving relative to each other, over time. The research formalizes something practitioners in hiring and customer experience have known intuitively: that whether a conversation goes well depends less on any single exchange and more on the emotional trajectory over the whole arc.

For detection to work, you need to capture both the “when” (the temporal dimension) and the “who is influencing whom” (the relational dimension). TRACE builds both in explicitly, and the results suggest the design is right.

---

④ Three business scenarios where this matters

If you could score emotional entrainment automatically from a recorded conversation, what would you do with it? Here are three concrete scenarios.

Interviewer selection and development. Organizations spend significant resources on interviewer training, but measuring interview quality is often reduced to structured scoring rubrics that miss the interpersonal dynamics. An entrainment score derived from interview recordings would give a more direct signal: did the interviewer’s emotional presence help the candidate open up? Which interviewers consistently move candidates from guarded to engaged? That’s a different and arguably more meaningful dimension of interviewer effectiveness.

Call center quality management. Current QA processes rely on selective call monitoring, keyword flagging, and post-call surveys — all of which are noisy and delayed. Automatic entrainment scoring applied to every recorded call would allow teams to identify which operators are most effective at de-escalating negative emotional states and which are most likely to get pulled into the customer’s frustration. That’s the kind of behavioral data that makes training programs specific rather than generic.

Counseling and mental health support AI. Tracking emotional synchrony between a client and a counselor over a session — and across sessions — could offer a supplementary signal about session quality and therapeutic alliance. Moments when emotional distance appears could prompt reflection or supportive feedback for the counselor, rather than relying entirely on self-report.

---

⑤ What to think through before building on this

Emotional AI systems of this kind come with real considerations that need to be worked through before deployment.

Privacy and informed consent. Analyzing the emotional content of recorded conversations is not a small thing. Interview candidates, call center customers, and counseling clients are in inherently unequal positions relative to the organizations recording them. Collecting emotional entrainment data from these interactions requires clear disclosure, genuine consent, and governance policies that specify how the data can and cannot be used.

Interpretability. A low entrainment score doesn’t mean the interviewer did something wrong. Cultural differences in conversational norms, the emotional weight of the topic being discussed, and individual personality differences all shape how emotional synchrony manifests — or doesn’t — in a given conversation. Scores are signals to investigate, not verdicts to act on.

Scope of current research. This study works from audio only. Visual cues, body language, and other modalities are not included. The dataset is controlled and synthetic in part. How well the approach generalizes to the full diversity of real-world conversational contexts will require further validation before organizations should treat the outputs as reliable at scale.

---

Conclusion

What this research establishes is that emotional entrainment — long understood as an important but hard-to-measure property of human conversation — can be detected from audio with high accuracy, provided the detection system accounts for both the temporal flow of the conversation and the relational dynamics between speakers.

For organizations that conduct high-stakes conversations at scale — in hiring, customer support, or mental health — this matters. The qualities that distinguish an effective interviewer from a mediocre one, or an exceptional operator from an adequate one, have historically been assessed through observation and intuition. A reliable automatic signal for emotional synchrony would make those assessments more systematic and more actionable.

The work also points to a broader principle for emotional AI design. Emotion is not a static attribute of a speaker at a point in time. It is relational and temporal. Systems that model it as either a snapshot or an individual property will miss what’s actually happening. TRACE is a contribution toward getting the architecture right.

---

Reference

1. Sathvik Manikantan Napa Ugandhar, Hao Zhang, Alison Gunzler, Yuzhe Wang, Thomas Thebaud, Georgi Tinchev, Venkatesh Ravichandran, Laureano Moro-Velázquez (2026). TRACE: Temporal Relationship-Aware Conversational Entrainment Detection in Dyadic Speech. arXiv preprint (2026).

* This article was written in part with AI assistance and may contain inaccuracies.