A US prison was recently able to detect and prevent inmates from carrying out illegal business by using artificial intelligence to analyse calls made into and out of the prison for unusual patterns.

All calls that connect on prison phone lines in the US are recorded and monitored, but it is a very time-consuming and boring job to listen endlessly to thousands of hours of conversations between inmates and their families in order to try to work out if anything nefarious is going on.

To speed the job up, one prison in the American mid-west is using a machine learning system designed by London-based firm Intelligent Voice to listen to all the audio files and detect odd patterns for humans to take a closer look at, according to New Scientist.

Preventing prisoners from carrying out illegal business

When the prison wardens looked at the analysis report created by the software, they discovered that one of the most popular non-trivial phrases being used by inmates was the phrase "three way".

Initially the wardens thought that the phrase was some sort of sexual reference, but because the word was used so often in phone calls they investigated further, and realised that the phrase was actually a secret code word.

All inmates are only allowed to call a few people on a previously agreed list of numbers. So if the prisoner wants to deal with illegal business and call someone else, for example, gang members on the outside, they need to call a friend or family member on the agreed list of numbers and then ask that person for a "three way", ie to dial the person they actually wanted to speak to into the call.

How a computer can transcribe audio better than a human

Intelligent Voice works by indexing key words and phrases from phone calls, so that prison wardens can then search for keywords in the telephone calls as if they were text, instead of audio. The system includes an Nvidia GPU, which processes all the calls at speeds of up to 400 times faster than real time, while additional modules analyse the calls to pick out specific popular keywords.

Nigel Cannings, CEO of Intelligent Voice, said that he invented the solution through taking machine learning algorithms and seeing what would happen if he were to get the computer analyse the audio waveform pattern of spikes and troughs instead of listening to the audio recording directly.

Cannings and his team of developers quickly realised that they could gain new insights from audio files by analysing data using techniques traditionally reserved for classifying images, and from the first tests, the solution had 88% accuracy.

The Intelligent Voice machine learning system is so clever that it taught itself to transcribe speech by analysing the inmates' phone conversations and comparing them to recordings of US congressional hearings, and the end result is accurate and much cheaper than getting a human to sit and manually transcribe conversations from audio files.