Oxobox’s platform extracts the audio track from a proxy file and moves it to the Lyra environment, where an inference algorithm analyzes the file to detect instances of human voice. Lyra’s machine learning system has been trained using thousands of segments from subtitled and QCed series and feature films by experts, allowing it to accurately identify human voice and generate a perfectly-timed subtitle file without text.

Subsequently, this file is adjusted through a process that follows the Netflix Specs Guide for Timed Text. This post-process can be customized to produce different results depending on the type of content (for instance, children’s content requires slower reading speeds in the subtitles).

In just 20 seconds, Lyra can generate an SRT file for a 45-minute episode, fully complying with Netflix’s specifications regarding duration and minimum gaps between lines. This process allows us to focus talent recruitment more on the linguistic area than on the technical one, resulting in improved quality and optimized costs.

Moreover, with this workflow, we can determine from day one, with absolute precision, how many subtitle and dubbing lines a project will need, contributing to better resource estimation and more accurate planning for large-scale content localization projects.