Skip to content
BleepsterAdd to Chrome

Product technology

How Bleepster’s AI YouTube Profanity Filter Works

See how Bleepster combines YouTube captions, context-aware AI, word-level timing, local audio control, result caching, and honest protection status.

By lirik
Bleepster turns a YouTube caption track into timed mute instructions. It reads caption text, asks a context-aware AI model to identify profanity, aligns the result to caption timing, and changes only the local tab audio during playback.

The filtering pipeline

  1. Caption retrieval. Bleepster obtains an available YouTube caption or transcript track and its timing information.
  2. Context-aware analysis. Caption text is sent over HTTPS through an authenticated Bleepster server function to Qwen on Alibaba Cloud Model Studio.
  3. Word-level alignment. Detected spans are mapped back to the timing available in the caption track.
  4. Local playback control. The extension lowers audio in the YouTube tab and optionally adds a replacement bleep sound.
  5. Local caching. Analysis results can be reused for the same video, so a rewatch does not require another Free-plan analysis.

Why context-aware AI matters

A fixed word list cannot reliably distinguish every use, spelling variation, inflection, transliteration, or mixed-language phrase. Bleepster evaluates caption text in context, which lets one system work across languages without asking the user to maintain a long dictionary.

AI analysis is included with Bleepster. Users do not supply a personal API key and do not need a separate account with an AI provider.

What stays on the device

Audio muting and the optional replacement sound happen locally in the browser. Bleepster does not download, record, store, or upload video audio. Settings, the local result cache, and the salted cryptographic hash of an optional Parent PIN are stored in the Chrome profile.

Why the protection status can change

The extension distinguishes a protected video from a video it could not analyze. Missing captions, unusable timing, a network error, or an unavailable analysis service can stop the pipeline. In those cases, an honest status is more useful than a silent failure.