I used local AI to fix bad subtitles in my media library

The reporter enjoys watching old-school movies during work and notices that many of these films lack accurate subtitles. Over the years, the reporter has collected various media sources, but even when subtitles are available, their quality varies greatly. The reporter has tried tools like Bazarr but found limited success. Recently, while reorganizing the media library, the reporter discovered Crowbarr, an open-source tool that utilizes local AI to check, repair, and generate subtitles for media libraries.

The tool is particularly effective in handling older media, such as 90-year-old horror movies, where subtitles often have poor timing or mismatched text.

Crowbarr runs on a NAS with Plex for playback across devices and the popular arrr stack for organization. While Bazarr excels at finding subtitles for modern content, its accuracy diminishes with older media. Crowbarr addresses this issue by comparing subtitles with actual dialogue using Whisper, an open-source speech recognition model. If the subtitles are accurate, they are left unchanged; otherwise, Crowbarr adjusts timing or generates new subtitles.

One significant advantage of Crowbarr is its ability to handle the absence of subtitles altogether. The tool integrates deeply with Radarr and Sonarr, automatically analyzing and verifying media and subtitles upon import or upgrade. Installing Crowbarr is straightforward using the provided Docker Compose file, and users can enter their Radarr, Sonarr, and Plex API keys or directly import media files.

The tool runs speech recognition locally, which may require more powerful hardware, but it supports GPU acceleration for faster processing. Crowbarr generally performs well with English media, although accents and spelling errors can occur. Overall, Crowbarr has proven to be a valuable addition to the reporter's self-hosted media stack, solving subtitle issues for a wide range of content.