Talk:How to write a new method of commercial detection
This may be a locale issue. With Cox Communications (Cable) in the US there is *always* a little black box at the beginning of every program segment in the upper left hand screen that shows that rating of the show. I'd imagine that'd be pretty easy to pick up and detect.
no idea how to use a wiki, sorry for formatting etc
Idea: Comercials are ALWAYS louder than the Program that they are in. you are sitting there with the volume just right, suddenly you have a coke comercial waking the neighbours. someone should analize just how much extra bass is in comercials and how little silence and add this to comercial flagging.
Addition - by Bg - I believe that commercials are recorded to broadcast with all the frequency bands (as per graphic equaliser) maxed out compared with broadcast programs where the frequency bands are adjusted to give proper balance and appropriate affects for that programme, this gives the impression of a volume shift (i.e. a global equal increase across all frequency bands)....
Addidion - by b00 - That's right, they're compressed to be constantly as close to 0dB as possible. So that means that there's a pattern of scenes with under-1dB dynamic range or so with very short pauses in between. This is what the commercial producers do, not the broadcast stations. The former just wants to be loud, the latter wants things to stay realistic. Neither can cross 0dB.
Idea: - by forge One way to sort those grays is to remember that commercials are repeated. So when myth is unsure weather to flag a 30 second block as a commercials or not just look for that block being repeated elsware among the recorded shows. If you want to really get clever, check if the block before and after it are the same in the secondary copys.
Properly implemented that could weed out most of what's left. Also implementing this should open the door to do something I like. A "novelty pool" Where you can choose watch all those commercials that were recorded for the first time today (or this week etc..) It has been my experience that sometimes a commercial is worth watching because 1. It could be humerus, 2. It mentions a product which is of interest to you.
PS: Maybe I just never studied Myth enough to know this is impossible.
Ideas: - by sgn How about text detection - ads always have some kind of text (higher frequency than shows) whether it is telephone numbers or company name.
Detection of known company logos might be useful also.
How about running the audio through a speech to text converter, and then the resulting text into a spam filter?!
Idea: - by aspect Many of my recorded movies could be separated from the commercials by the different picture aspect. Movie part ist often widescreen eg. 16:9 where the commercials are still often 4:3
An algorithm for monitoring the percieved loudness..
As several people have suggested, the sound on commercials is often perceived as louder, even though some sources suggest that the actual RMS and/or peak volumes are in fact the same or lower than the programme content. Thus one obstacle in using loudness for commercial detection is acurately measuring the perceived loudness. The ReplayGain algorithm is one approach to this problem.
The ReplayGain algorithm was originally intended for altering the volume of MP3s from various sources have so that they have similar perceived loudness. See http://replaygain.hydrogenaudio.org/calculating_rg.html for details.
Perhaps one could apply this algorithm over a rolling window and look for the loud sections. Even better - the statistical spread of RMS energy values is probably quite different for commercials and programme content - (See )