Regular quality control of Colombian cocoa liquor (CL) has been constrained by the limited availability of expert sensory panels and the absence of a reconciliation framework linking instrumental and human sensory data. This study developed a multimodal sensory dataset comprising 50 CL samples by integrating human sensory evaluation (HSE) from 187 panelists with electronic sensory analysis (ESA) using an electronic tongue (ET) and electronic nose (EN). Instrumental signals were transformed into sensory descriptors: ET signals were converted into taste deviation scores via the Approaching Target method relative to a balanced CL reference, while EN time-series data underwent sensor reliability filtering, stationarity testing, and cross-correlation with categorized aroma standards. Both modalities were mapped onto a shared 29-descriptor vocabulary and z-normalized for direct cross-modality comparison. Gaussian Mixture Model (GMM) clustering using Bayesian Information Criterion selected a 12-component tied-covariance structure. Cross-modality alignment yielded a Normalized Mutual Information (NMI) of 0.554, indicating that electronic sensing and human perception provide complementary rather than redundant representations of the cocoa sensory space. ESA captured broad chemical signals including sub-threshold volatiles, while HSE reflected perceptually salient flavor characteristics. The curated multimodal dataset may support future supervised learning studies on attribute classification or quality prediction, but predictive modeling falls outside the scope of the present sensory-characterization study. • Fifty samples of cocoa liquor from various associations in Colombia were analyzed through both electronic analysis and human sensory evaluation. • Comprehensive workflows transforming raw ET/EN sensor data into flavor and aroma descriptors that correspond directly with human panel intensities. • Sensor reliability analysis was conducted to verify that the electronic equipment’s response is dependable and suitable for comparison with human evaluation. • Non-parametric tests and PERMANOVA indicate substantial panel training effects on attribute detection and intensity assessments. • BIC-selected twelve components and constrained covariance for Gaussian Mixture Models produce complementary macro-clusters across electronic and human modalities. • Case studies illustrate the effective detection of sub-threshold volatiles and the enhancement of impact flavors through panel assessment, facilitating machine-learning forecasts of sensory quality. • The ability of electronic sensory evaluation to identify most attributes in comparison to human perception and produce objective assessments. • A detailed data frame of electronic sensory analysis and human sensory evaluation for 50 distinct cocoa liquor samples was generated for subsequent analysis and application.
Quintana-Rojas et al. (Fri,) studied this question.