I watched an interesting video by Ben Eater where he takes apart the TV Guardian to extract out the curse words installed in the EPROM of the device. It’s delightful to watch him use the spreadsheet to easily reverse engineer the logic of censoring. However, in the video it’s not apparent what TV Guardian does to achieve the task at hand.
This made me realize I have never given much thought to the inner workings of beeps for the censored words in TV shows. This post is just me thinking out loudly about how I would have attempted to deliver such a feature for the TV back in the DTH days.
To begin with, let’s see how to go about beeping out the words in movies in a very 101 manner. Assuming you have subtitles synced perfectly with the audio tracks, then you could just have simple rules to catch the words matching against the target word list. And, for every match overdub the audio track with beep sounds, keeping safe pre & post margins. Additionally, replace the word in the subtitles with the safe word, as shown in the above mentioned video by Ben.
This approach has obvious flaws:
- It cannot work without subtitles or closed captioning.
- Difficult to beep out perfectly given variance in the usual speech rate
Digging around, I found that TV Guardian’s website claims that even if there is no caption visible on the TV, it is required by law to have hidden closed captioning for movies and TV shows (I am not sure about other countries in this context). If this is the case, our 101 solution can work to clumsily address the issue at hand.
And, to address situations where matters are near real time (live TV), I found out that Broadcasting networks usually keep a profanity delay with broadcast operators ready to press the mute or beep button if it comes to that. However, I believe most of us have heard curse words on “live” TV which were not beeped — Careless operator, eh?
I haven’t extensively researched the implementation of ML models for Live TV profanity filtering yet, but I know OpenAI has content filtering API to solve similar problems. If I were to make such a system I would utilize similar APIs to create a system which could assist the operator reduce such misses. I believe that it can be an interesting project to implement. Maybe I will do that some other time.