GSM is actually already doing half the processing for you by reducing the dimensionality of the data. Pattern-matching a few dozen words using a custom neural network ASIC could probably be done in real time with very little hardware.
The calls are cut off because it's simple and effective. If people can communicate and organize, the damage is already done and you have to respond with force to stop protests. But if you can prevent protests from ever being organized in the first place...
>GSM is actually already doing half the processing for you by reducing the dimensionality of the data
Could you expand on this? Or just give me some key terms that I can look up to get more info. Sounds very interesting.
EDIT: Also, I don't buy the argument that dropping the calls is simple and effective. The antigovernment types are simply going to start using trivial codewords, like every criminalised person throughout history has done. Therefore it will barely touch them.
However it will drive businesses, both chinese and foreign, crazy. They are not interested in codewords. If you criminalise free speech, only criminals will speak freely.
Lossy compression involves throwing out redundant data; if you look at the data as a vector, you are effectively reducing the number of dimensions in the vector. Speech recognition involves rearranging the data to find its primary components (e.g. pitch, timbre, etc.). Most voice / audio compression schemes perform a similar rearrangement, as this is a convenient way to find (and eliminate) redundant data.
In particular, Fourier transforms, wavelet transforms, cepstrum transforms, subband filtering, and principal component analysis are techniques common to both audio compression and speech recognition.
Comments
GSM is actually already doing half the processing for you by reducing the dimensionality of the data. Pattern-matching a few dozen words using a custom neural network ASIC could probably be done in real time with very little hardware.
The calls are cut off because it's simple and effective. If people can communicate and organize, the damage is already done and you have to respond with force to stop protests. But if you can prevent protests from ever being organized in the first place...
>GSM is actually already doing half the processing for you by reducing the dimensionality of the data
Could you expand on this? Or just give me some key terms that I can look up to get more info. Sounds very interesting.
EDIT: Also, I don't buy the argument that dropping the calls is simple and effective. The antigovernment types are simply going to start using trivial codewords, like every criminalised person throughout history has done. Therefore it will barely touch them.
However it will drive businesses, both chinese and foreign, crazy. They are not interested in codewords. If you criminalise free speech, only criminals will speak freely.
Lossy compression involves throwing out redundant data; if you look at the data as a vector, you are effectively reducing the number of dimensions in the vector. Speech recognition involves rearranging the data to find its primary components (e.g. pitch, timbre, etc.). Most voice / audio compression schemes perform a similar rearrangement, as this is a convenient way to find (and eliminate) redundant data.
In particular, Fourier transforms, wavelet transforms, cepstrum transforms, subband filtering, and principal component analysis are techniques common to both audio compression and speech recognition.