How was 'Adblock Radio' designed to skip web radio commercials designed?

by Eric Nopanen

" Adblock Radio " is a service that can skip advertisements (CM) inserted in web radio programs, like the extension of web browsers blocking advertisements on websites, was born at the end of 2015, 2019 As of year, it corresponds to more than 60 radios in 7 countries. On the contrary, Alexandre Streli, the producer, has revealed how this "Adblock Radio" was designed.

Adblock Radio
https://www.adblockradio.com/

GitHub - adblockradio / adblockradio: An adblocker for live radio streams and podcasts. Machine learning meets Shazam.
https://github.com/adblockradio/adblockradio

Designing an audio adblocker for radio and podcasts
https://www.adblockradio.com/blog/2018/11/15/designing-audio-ad-block-radio-podcast/

Mr. Strel started "Adblock Radio" at the end of 2015, a few months after completing the doctoral course of nuclear fusion plasma physics. When popularity comes out of service in 2016, he says that he received pressure from a French radio station through a lawyer.

In order to accurately find out the CM portion from the radio program being broadcast, it is necessary to detect and distinguish where the main part is and which is CM.

First thing Mr. Strel said was "to detect it by the difference in volume." If it is a classical music-centered program, since the sound of the CM part is conspicuous, it was effective, but in the pops the difference between the volume with the CM is small, abandoning.

The method of "Block by specifying the time when CM is inserted" fails because the time of CM flow is not fixed. The method of "using metadata of web radio program" is also discouraged because it is not useful for identification. Finally, with the help of a person, I arrived at "Where the listener got a button to be pressed at the start and end of CM", but in the newly added stations I got enough listeners for the function to work properly There was a fear that the user would leave because the bootstrap did not work well.

Mr. Strei, who twisted as if he could handle it automatically, then started "voice recognition and lexical analysis". However, Mr. Strel had been an amateur with respect to voice processing, and it was impossible to create a system that successfully recognized CM. Since then, since open source tools such as MozillaDeep Speech have appeared, it is worth reviewing if it is a 2019 level technology, Mr. Strei said.

Through trial and error it is "Detection by acoustic fingerprint and database of CM". Acoustic fingerprints are converting speech into a series of numbers. If the acoustic fingerprint that appears in the sound source being broadcast matches with the sample in the CM database in many places, it can be judged that the CM is being broadcasted exactly. At this time, although there was erroneous detection to miss the CM, there was almost nothing to erroneously recognize the main part of the program as CM. Regarding the problem that the CM being broadcasted will be updated to a new one, record a radio program and list the sounds that flow as many times as everyday. We cleared and extracted only the data of about 30 seconds which is the length of the general CM.

In this way, "Adblock Radio" which has been improved and further improved is said to have legal and technological problems to execute in the cloud, but it is said that "I want to advance integration with radio players" "In the future, audio advertisements will be in memories," he says.