- 0 Comments
If you have never heard of an exabyte, don’t worry, you are not alone. One exabyte is equal to one billion gigabytes and, until a few years ago, the word had no practical application. Now, however, however litigators need to think about how they are going find relevant data in such massive quantities of data. According to IT analyst firm IDC, 161 exabytes of digital information was created last year, a figure that will rise to 988 exabytes by the end of the decade. No individual corporation has that amount of data currently, but they are headed in that direction. Exxon Mobil has more than two petabytes (2,000,000 gigabytes) of operational data on line, and much more in storage. Chevron adds two terabytes (2000 gigabytes) to its data stores every day.
While organizations have long had tools to do text searches to locate discoverable data, text documents are not necessarily the principal drivers behind exploding storage needs. A large part of this growth is due to increasing amounts of non-textual data such as photos, videos and voice recordings. However, by taking the latest voice recognition technology, and joining to a search engine, it is now becoming as easy to locate a relevant voice mail as to find an email.
Sources of Data
When it comes to producing or discovering voice data, the first action, as with text documents, is determining that it exists, where it exists and in what format. For individuals, this includes voice mails, personal videos, videoblogs, pod casts, and postings on sites such as YouTube or MySpace.
A YouTube or MySpace posting showing the plaintiff dancing at a friend’s party can undercut a personal injury claim without paying a PI to follow and film the person.
Corporations similarly are creating vast quantities of spoken digital data which is now recoverable. Most companies, of course, use voice mail, but it goes far beyond that. For corporate or securities litigation, in addition to looking at SEC filings and press releases concerning earnings, corporations are also releasing data in other formats. If you go to Wal-Mart’s website, you can view a four-hour video presentation of the June 1, 2007, annual shareholders meeting, and General Motors posted the webcast of its global automotive security analysts conference.
Medical records also increasingly incorporate spoken data into the patient records. Miami Children’s Hospital is conducting a trial with IBM to enter surgical data into the patient record verbally. “A lot of environments in the hospital are hands free and some are eyes free,” says Redmond Burke, M.D., Chief of Cardiovascular Surgery at Miami Children’s Hospital. “When I am looking at a baby’s heart, I can’t look up at the monitor or enter data into the patient database.”
With this system, microphones and speakers in the operating room are connected to the hospital’s patient records system. At the start of an operation, the surgeon issues a verbal command to access the clinical database, and the patient data is read over a speaker system in the operating room using voice synthesis. The voices of the operating room personnel, together with instrument readings and photos taken during the procedure, are then recorded back into the patient’s medical record.
To improve collaboration of geographically distributed work forces, businesses are using virtual office applications IBM’s Lotus Sametime, EMC Corporation’s eRoom or Viack Corporation’s VIA3 to allow distributed personnel to collaborate on projects through document sharing, teleconferences and videoconferences. These conferences are often stored in the virtual office space for later reference.
Companies are also switching from separate voice and email systems to Unified Messaging Systems which combine email, voice mail and faxes. Users can receive voice mails in their email in-box, and click on them to listen to the message.
Some of the biggest voice storage systems are being used by customer call center operations. Those messages that say that “calls may be monitored or recorded” don’t necessarily mean that someone will be listening to a tape recording of a conversation, but that it will be captured and analyzed by a call center software such as Autonomy’s eTalk. “The systems can listen to the conversations in real time and put suggested answers up on the agent’s screen,” says Mike Lynch, Autonomy Corporation’s founder and CEO.
“You record all these calls and cluster them by meanings, so your marketing and engineering staff can tell what customers are calling about and take appropriate actions.” The software can also detect when a customer is starting to get angry and signal a supervisor to take over the call. The phone data is stored for analysis such as what types of issues are commonly being mentioned by end users. For product liability litigation purposes, these records can show, for example, when a company first started receiving complaints about one of their products, how many complaints were received, and what actions the company took to address its customers’ concerns.
Then there are all the millions of public and private security cameras recording images around the clock. The University of Pennsylvania, for example, has more than 400 watching its campus, but that is nothing compared to the estimated half million peppering London’s streets.
Lit’s a Beach
Given the sheer quantity of voice and video data now available, the problem is how to find the relevant pieces for production. You don’t want to have someone sit there and listen to everything to see if any of it pertains to the issues at hand. Just as there are tools for searching written data, similar tools now exist for voice data.
Searching voice data is a two-step process. The first is to break that continuous stream of sound down into phonetic units. This can be done with a computer’s regular processor or with a plug-in board specifically designed for the task. Natural Speech Communication Limited, for example, makes boards that can analyze up to 120 simultaneous conversations and issue an alert when a key word or phrase is said.
The phonetics are then converted into words and a transcript can be generated. That can be done immediately, but the safer path is to keep the original recording or phonetic data and then do searches against the phonetics rather than the transcript. As anyone who has experience with voice recognition software knows, the technology is not perfect.
Voice recognition software works by making a series of guesses about what the sounds mean, just as humans do. As the sentence or conversation progresses, we adjust the meanings we assign to words based on context. Lynch gives the example of the phrases “recognized speech” and “wrecked a nice beach.” Those phrases are almost identical phonetically and only by knowing context can the meaning be elicited. If the words “tanker,” “oil” and “spill” are also part of the context, people and voice recognition software recognize that “beach” is the correct word. Further complicating the issue of creating an accurate transcript are the speakers’ accents.
“It could be a Texan talking about country music or a well-spoken Brit speaking about Tony Blair’s government,” says Suranga Chandratillake, co-founder and CTO of blinkx, a video search engine that currently indexes more than 12 million hours of web video content. “It has to understand what was said, regardless of the accent, so you need as many contextual clues as possible to determine the meaning.”
Then there are the problems of distinguishing the voices from other noises present. For example, the operating room system was adjusted to filter out the sounds of the operating room equipment and only respond to human voices. Call center systems are designed to filter out cell phone static and background noises. “
Voice technology continues to be an evolution,” says Rich Cox, the vice president for AT&T Laboratories who oversees voice research. “Speech recognition is never perfect, but people are learning what they can do with it and what to do to work around its shortcomings.”
Since companies are now recording much of this voice data as part of their business processes, it is discoverable. When crafting a litigation hold, therefore, counsel must look at all the types of voice records that may be relevant to the case and ensure they are not destroyed.
“You need to think about what actions people took that help tell the story of what really happened – sent email, wrote memos, talked on the telephone, went and did something at a particular place – that gets you thinking about what technologies are going to be involved,” says Conrad Jacoby, an attorney and founder of Efficient EDD, an independent consultancy specializing in e-discovery and litigation management. “If a substantive part of the story involved conference calls or a video conference, then we need to say what are the tools that people used to take this action and preserve those records.”
At that point, counsel should be working closely with IT staff to ensure the litigation hold is properly crafted and implemented. But Jacoby says there is no need at that early stage to call in an expert, unless the legal team is too busy or is unable to understand the IT systems. This should all be done before the initial discovery conference.
Next comes the task of determining what is actually relevant. In this case there are two tools – Nexidia Forensic Search and Autonomy’s IDOL (Intelligent Data Operating Layer) search technology which come into use. Nexidia’s product is directly geared toward searching audio and is available as software or as a service. Autonomy searches both documents and audio files.
Either of these products can be used to create a transcript of the audio files and conduct key word search or concept searches on, for example, a company’s voice mail system. Given that the speech recognition technology is not 100% accurate, the search will produce some unnecessary audio files, so they will have to be listened to to ensure they are relevant. The search is also likely to miss some relevant files.
“In that key cell phone message with a loud band playing in the background, it is almost impossible to recognize the voice and so will fall out of any of these filtering paradigms,” says Jacoby.
Nevertheless, using these search tools makes it possible to discover voice data that would be impractical to find if it meant hiring an army of assistants to listen to and transcribe every voice mail. But since different voice recognition software and search parameters will turn up different data, it is important that the search criteria are agreed upon by both parties.
“Both sides may agree to use the same technology so neither has an edge,” says Jacoby. “Your results will have the same problems as my results, so we are on even ground.”