As discussed further below in relation to FIGS. 5 and 6, correlations may also be determined between feature sets generated from video frames compared to predicted feature sets generated from an audio bin (FIG. 5) and correlations may be determined from feature sets generated from audio bins compared to predicted feature sets generated from a video frame (FIG. 6). In some implementations, multiple correlations may be determined for an audio bin/video frame pair based on the various techniques discussed herein, e.g., FIGS. 5 and 6.

The video component and the audio component may be determined not to be synchronized based on the correlations (308). In some implementations, the correlations are provided as inputs to a model, such as a classifier. The classifier may be trained on correlations between video components and audio components of media presentations from a training data set. In some implementations the correlations may include correlations between each audio bin of a portion of the audio component and each video frame of a portion of the video component of a media presentation. The model may analyze the correlations to determine whether the audio component and the video component are synchronized. In some implementations the model may implement a threshold strategy, where an audio bin and a video frame sharing a media timeline are considered desynchronized if the correlation does not exceed a threshold value.