Trending Topics

The potential for using machine learning for reviewing inmate communications

These digital tools have the potential to determine threats and even provide insights into the behavior and needs of inmates


With emerging technologies like machine learning, the power to analyze and interpret vast amounts of inmate communication data is now at our fingertips.

Getty Images

By Christopher Ditto

Average inmate call prices in the U.S. are trending downward and inmates are making more phone calls than ever.

From a sampling of internal numbers offered by major call providers, it is expected that inmates will spend at least 12 billion minutes speaking on the phone this year, and most of those calls will be digitally recorded. These calls are, for the most part, of no consequence to anyone other than the participants, capturing everyday conversations between inmates and loved ones they miss, or with friends who might be helping ensure the pets are fed, or the power bill is paid.

But every once in a while, portions of these inmate conversations can present a security threat to inmates, correctional staff and the public, creating an acute need within correctional facilities to correctly identify and stop these threats. With emerging technologies like machine learning, however, the power to analyze and interpret vast amounts of inmate communication data is now at our fingertips. And these digital tools have the potential to determine threats, and even provide insights into the behavior and needs of inmates.

Flagging security threats

In recent years, the use of machine learning technology has become increasingly prevalent across a wide array of industries, and criminal justice is no exception. Combining machine learning processes with some human review can allow correctional facilities to review all inmate communications rather than simply listening to just a small fraction, and make data-driven decisions to bolster organizational security and enterprise operations.

Inmates can engage in various illegal activities within correctional facilities, making it all the more important that staff members have the tools in place to mitigate threats. One of the most commonly reported activities is extortion, in which inmates pressure their families to fund other inmates or directly threaten people outside. Additionally, inmates can engage in conspiracy by coordinating with outsiders to hide evidence or alter testimony. Inmates may also threaten facility security by smuggling contraband or planning escapes that involve violence against correctional staff.

Currently, flagging these security threats in a mountain of call recordings is a massive challenge for correctional organizations. If all correctional phone recordings across the United States were to be reviewed by humans at a normal speed, it would take a team of 100,000 people, working full-time, just to keep up.

To alleviate that logistical hurdle, the traditional strategy adopted by correctional facilities is to target a specific percent of calls to review and then to increase the odds of finding actionable intelligence by focusing on calls from inmates with violence or drug-related charges, or gang affiliations, or prior convictions.

As a strategy, most would agree that searching for a needle in a haystack, while ignoring the vast majority of the haystack, isn’t the best option, but for correctional organizations who cannot do more than listen to calls, this has long been the only viable approach.

But machine learning can provide robust improvements to moribund analog processes. This technology is not some faraway possibility but a reality that is already changing how corrections facilities review inmate communications.

How machine learning can assist

Machine learning can assist in reviewing inmate communications by automatically transcribing and translating audio, scanning for watchwords and phrases, creating word clouds and conducting natural language processing, analyzing patterns, and providing spot audio functionality. It can also convert inmate calls into text with high accuracy, scan for keywords and names, create visual representations of conversation topics, and analyze communication patterns to identify suspicious activity.

Additionally, by associating each word with a timecode, call monitors can skip to specific parts of the call recording to hear words in context. These features can improve the ability of algorithms to understand and interpret the meaning of communications and help flag suspicious activity or security threats for human review.

A sample process using these techniques might begin by automatically transcribing all calls, detecting and translating foreign languages, and searching the recordings for watchwords, suspicious phrases, signs of three-way calling (where the call recipient bridges in a third party, a practice generally prohibited at correctional facilities), and suspicious meta-data (calls and deposit patterns that align closely with past actionable calls). These initial automated processes can flag calls that have a higher likelihood of being of interest to investigators.

For the next level review, software can take calls flagged with the first step, extract interesting words or phrases, create word clouds, or process the call transcripts into a readable summary, just as a human might summarize a call if given enough time. Human reviewers can review word clouds or read transcript summaries very quickly, and determine if any calls deserve additional attention. For the smaller percentage of calls that do, reviewers can click keywords or phrases to listen to the section of the call where that content was detected, allowing reviewers to hear the selected text spoken in context. For example, a phrase like, “I’m going to kill him” may imply something very different if said in a jestful or friendly tone than it does when delivered flatly. As a last step, calls that still haven’t been dismissed by the review processes so far, can be reviewed in their entirety, either through a full reading of the call’s transcript, or with a full listen of the call recording.

With the help of machine learning, correctional facilities can be better prepared to deliver positive outcomes for incarcerated populations and staff members, and improve inmate experiences while promoting successful reintegration within communities on the other side. What might have once been in the realm of science fiction is something that can now create tangible change for countless individuals and communities within the criminal justice system; that momentum will only grow as the technology becomes more powerful and precise.

About the author

Christopher Ditto is the Vice President of Research & Development for ViaPath Technologies, a provider of inmate communication technology in the United States. Over the past decade, he has worked on building inmate communication and tablet resources, implementing technology for over 1,000 correctional facilities serving over 800,000 inmates daily, as an engineer, software architect and project manager.