Techniques are described for generating metric(s) that predict survey score(s) for a service session. Model(s) may be trained, through supervised or unsupervised machine learning, using training data from previous service sessions between service representative(s) and individual(s). Training data may include, for previous service session(s), a session record (e.g., audio record) of the session and a set of survey scores provided by the serviced individual to rate the session on one or more criteria (e.g., survey questions). The model(s) may be trained to output, based on an input session record, metric(s) that each correspond to a survey score that would have been provided by the individual had they completed the survey. The model may be a concatenated model that is a combination of a language model output from a language classifier recurrent neural network, and an acoustic model output from an acoustic feature layer convolutional neural network.