Purpose <p>Cocaine is one of the most widely abused drugs, and the analysis of seizures including drug profiling is a critical task for forensic practitioners. Street cocaine samples frequently contain additives, and they could hinder conventional drug analysis. In this study, a multivariate analysis method was made to rapidly estimate the composition and quantity of cocaine and its additives using Fourier transformation-infrared absorption (IR) spectroscopy and Raman spectroscopy.</p> Methods <p>IR and Raman spectra of standard cocaine samples and 21 additive samples were acquired. Training dataset was synthetically constructed by making linear combinations of the standard spectra with randomly generated weight and noises. Four different models of non-negative least square regression (NNLS), non-negative matrix factorization (NMF), non-negative elastic net (Enet) and multilayer perceptron (MLP), were constructed. The constructed models were applied to street samples whose components were previously characterized by conventional methods.</p> Results <p>Though all four methods gave good root-mean-square errors on the training dataset, application to street samples gave different results. NNLS provided the sparsest and the most appropriate results. MLP also provided similar level prediction capability. NMF could not retrieve the original standard spectra. Enet could not give sparse results and application to the street samples showed large fitting residuals. Based on these results, a decision procedure to predict the contents of cocaine seizure was constructed by the combination of NNLS and MLP.</p> Conclusion <p>Vibrational spectroscopy combined with multivariate analysis demonstrated significant potential for the rapid, semi-quantitative prediction of cocaine seizures with the information on diluents/adulterants.</p>