- The paper introduces Earinter, a closed-loop system using commodity earbuds for real-time monitoring and intervention of eating pace, achieving a 36.72% reduction in food consumption rate during field tests.
- The system uses bone-conduction sensing and EfficientNet-B0 classifier with 97% F1 score for chewing detection and chewing per swallow (CPS) for effective intervention.
- The field study involving 14 participants shows a significant increase in chews per swallow (74.22%) under intervention, highlighting the potential for reducing eating speed.
Earinter is a closed-loop, single-device system that monitors eating pace in daily meals using a commodity earbud and delivers theory-grounded just-in-time (JIT) audio interventions when rapid eating is detected. The paper addresses a recognized gap in ingestive-behavior computing: while wearable sensing of chewing and swallowing has matured, most prior systems either confine support to post-meal reflection or rely on brittle fixed-schedule cues, with little coupling between real-time behavioral inference and intervention delivery policy. The work contributes an end-to-end pipeline spanning bone-conduction sensing, on-device inference, Wizard-of-Oz–refined intervention design, and a 13-day within-subject field evaluation.
Sensing and reasoning pipeline
The system repurposes the bone-conduction voice sensor in Honor Earbuds 3 Pro to capture chewing-induced vibrations in the ear canal, which are less sensitive to ambient noise than air-conducted microphones. Eating pace is operationalized as chews per swallow (CPS), chosen over chew-frequency metrics because CPS is directly actionable through deliberate mastication before swallowing.
The dataset comprises 10.07 hours of recordings (5.65 h in-lab, 4.42 h in the wild) from 18 participants, annotated against throat-microphone audio and synchronized video: 45,582 chewing events and 2,150 swallows in total. The pipeline has three stages. A lightweight state machine segments candidate chewing cycles from 50 ms frames via a 40 dB energy threshold, constrained to 100–400 ms durations. An EfficientNet-B0 classifier adapted from the EarVAS-Net framework, trained with Focal Loss for class imbalance, then performs binary chewing confirmation. Because commodity earbuds fail to capture swallow acoustics reliably, swallowing is inferred heuristically from inter-chew intervals: an interval exceeding 0.8 s, or 1.5× the running mean inter-chew interval since the last swallow, is classified as containing a swallow.
Performance supports closed-loop use: chewing detection achieves F1 = 0.97 (accuracy 96.86%, precision 99.02%, recall 95.05%) under cross-user testing; pace estimation yields MAE of 0.18 ± 0.13 chews/min at the meal level and 3.65 ± 3.86 chews per swallow. Deployed as ONNX on a smartphone, the full pipeline processes a 3-second window in 57.64 ± 24.89 ms, comfortably within real-time constraints. One limitation worth noting is that the fine-grained CPS error (MAE ≈ 3.7 chews/swallow) is non-trivial relative to typical CPS values around 15, though it proved adequate for threshold-based triggering.
Theory-grounded intervention design
Intervention design is framed by Dual Systems Theory: fast eating is treated as System 1 automatic behavior, and prompts are structured across the meal to re-engage System 2 processes — Goal Setting Theory pre-meal, brief micro-nudges during the meal, Self-efficacy-based summaries post-meal. Two Wizard-of-Oz pilots calibrated content and dose before automation:
- PS1 (content, N=6) found participants preferred concise, low-effort prompts; dense messages became annoying at high frequency, and bare chimes were ambiguous without explanation. Three prompt families were selected — System 1 micro-nudges, control-theory progress feedback ("you are N chews away from 25"), and gain-framed rationales — and expanded into varied-length libraries via GPT-4o.
- PS2 (frequency, N=10) showed monotonic annoyance increases with prompting density; the 30-second minimum interval scored lowest annoyance (mean 2.3/5, preferred by 6 of 10), versus 3.0 for 10 s. The final policy uses a 30 s cooldown with randomized sequencing of short prompts punctuated occasionally by medium or long ones.
Field study results
A 13-day within-subject study (N=14 screened fast eaters; CPS < 20 and intake > 50 g/min) compared a Control condition (pre-meal reminder only) against the full closed-loop Experiment condition (pre-meal reminder + JIT prompts + post-meal summary), each phase followed by a no-support retention day. Of 364 collected meals, 342 remained after excluding 6% for device issues or anomalous meals.
Chews per swallow rose substantially. Baseline CPS was 15.02 ± 3.67; the Experiment condition reached 26.42 ± 8.61 versus 17.88 ± 5.32 for Control. A negative-binomial GLMM estimated a significant condition effect (effect = 0.37, p < 0.0001), and crucially the difference persisted on retention days without any support (effect = 0.32, p < 0.001). CPS did not differ between active and retention phases of the Experiment condition (p ≈ 1).
Food-consumption speed dropped correspondingly. Linear mixed models showed the Experiment condition reduced consumption rate relative to Control by 10.48 g/min (EMM 47.82 vs. 58.31 g/min, p < 0.001), with persistence at retention (−9.29 g/min, p = 0.02). Baseline speed (76.40 g/min) was significantly higher than all post-baseline conditions, indicating that even the pre-meal reminder alone produced some effect, but the closed-loop intervention roughly tripled the reduction.
Subjective measures were consistent: Behavioral Positive ratings improved under intervention (β = 0.80, p < .001) and persisted; Fullness showed a strong odds ratio of 12.62 (p < .001); Comfort did not differ across conditions, indicating auditory prompting imposed no additional wearing burden beyond the earbud itself. Mean SUS was 70.0 ± 9.6 ("OK-Good").
Qualitative findings add an important nuance: although participants reported earlier fullness and slower pace, total food intake changed little because plate-clearing norms and fixed portions decoupled pace from intake for many users. This suggests the intervention reliably shifts chewing dynamics and satiety perception more than absolute consumption. Participants also favored actionable control-theory prompts over generic goal-framing messages, which some dismissed as common-sense, and requested personalization of targets to food texture and greater transparency about ongoing state.
Limitations
The authors are explicit about three constraints. The 13-day window cannot establish durable habit change despite significant retention-probe effects; longer deployments are needed to assess persistence and habituation. The uniform 25-chews-per-mouthful reference ignores food-dependence of healthy pace, so prompt thresholds are coarse across meal types. And the fixed cooldown policy may misalign with situational preferences, motivating future context-aware suppression. Methodologically, the qualitative analysis did not compute inter-rater reliability, and the sample of healthy young adults limits generalizability to clinical populations such as those with obesity or metabolic disorders.
Conclusion
Earinter demonstrates that a single commodity earbud can close the loop from sensing to theory-informed intervention for eating-pace regulation in the wild. The combination of reliable chew-level detection, heuristic swallow inference, WoZ-calibrated delivery policy, and statistically robust field results — including carryover onto unsupported days — constitutes a concrete instance of just-in-time adaptive intervention for habitual behavior. The open questions left by the paper concern long-term habituation, food-adaptive pacing goals, and whether pace improvements translate into reduced intake when portion norms do not constrain consumption.
