Busting “ghost subtitles” on streaming services

Abstract

In this paper we focus on a phenomenon that all subtitle users experience: “ghost subtitles”. “Ghost subtitles” are subtitles we notice in our peripheral vision, only to find them gone by the time our eyes have moved down to start reading or disappearing while we are still reading. “Ghost subtitles” often meet the minimum duration and maximum speed requirements set by platforms or broadcasters but disregard the time it takes to move gaze from the image to the subtitle (i.e., processing latency). The one-speed-fits-all approach means that, many subtitles are not on screen long enough to allow viewers to finish reading them, which could result in frustrating viewing experiences. To determine how prevalent fast subtitles are on streaming platforms, this paper presents an analysis of the distribution of subtitle speeds based on a corpus of subtitles from one of the major streaming platforms. We further investigated the impact of subtitle speed and audio language on processing latency based on eye-movement data from a total of 109 participants in two separate experiments. We found that almost 15% of subtitles in our corpus were faster than 20 cps, and almost 8% of subtitles were shorter than one second. We also found processing latencies of around 400 ms for fast subtitles to around 700 ms at speeds of 12 cps, and between 580 ms and 760 ms in different audio conditions. This points to the importance of setting subtitle speed and duration in a way that allows viewers enough time to process both the image and the subtitle properly.