speech analytics - the complete semantic index

8/3/2019 Speech Analytics - The Complete Semantic Index

1/8

The Complete Semantic Index

An Overview of Verints Speech Analytic Technology

April 2009

This document contains proprietary and confidential information of VerintSystems Inc. and may not be distributed to any persons or organizations for


2/8

Unauthorized use, duplication, or modification of this document in whole or in part without the written consent of VerintSystems Inc. is strictly prohibited.

By providing this document, Verint Systems Inc. is not making any representations regarding the correctness orcompleteness of its contents and reserves the right to alter this document at any time without notice.

Features listed in this document are subject to change. Please contact Verint for current product features and specifications.

All marks referenced herein with the or TM symbol are registered trademarks or trademarks of Verint Systems Inc. or itssubsidiaries. All rights reserved. All other marks are trademarks of their respective owners.

2009 Verint Systems Inc. All rights reserved worldwide.

2009 Verint Systems Inc. All Rights Reserved Worldwide.

Confidential and Proprietary Information of Verint Systems Inc.

All materials (regardless of form and including, without limitation, software applications, documentation, and any other

information relating to Verint Systems, its products or services) are the exclusive property of Verint Systems Inc. Only

expressly authorized individuals under obligations of confidentiality are permitted to review materials in this

document. By reviewing these materials, you agree to not disclose these materials to any third party unless

expressly authorized by Verint Systems, and to protect the materials as confidential and trade secret information. Any

unauthorized review, retransmission, dissemination or other use of these materials is strictly prohibited. If you are not

authorized to review these materials, please return these materials (and any copies) from where they were obtained.

All materials found herein are provided AS IS and without warranty of any kind.

The Verint Systems Inc. products are protected by one or more of the following U.S., European or International

Patents: USPN 5,659,768; USPN 5,790,798; USPN 6,278,978; USPN 6,370,574; USPN 6,404,857; USPN

6,510,220; USPN 6,757,361; USPN 6,782,093; USPN 6,952,732; USPN 6,959,405; USPN 7,047,296; USPN

7,149,788; USPN 7,155,399; USPN 7,203,285; USPN 6,959,078; USPN 6,724,887; USPN 7,216,162; EuropeanPatent 0 833 489; GB 2374249; and other provisional rights from one or more of the following Published US Patent

Applications: US 10/061,469; US 10/061,489; US 10/061,491; US 11/388,854; US 11/388,944; US 11/389,471; US

10/818,787; US 11/166,630; US 11/129,811; US 11/477,124; US 11/509,553; US 11/509,550; US 11/509,554; US

11/509,552; US 11/509,549; US 11/509,551; US 11/583,381; US 10/181,103; US 09/825,589; US 09/899,895; US

11/037,604; US 11/237,456; US 09/680,131; US 11/359,356; US 11/359,319; US 11/359,532; US 11/359,359; US

11/359,358; US 11/359,357; US 11/359,195; US 11/385,499; US 11/394,496; US 11/393,286; US 11/396,061; US

11/395,992; US 11/394,410; US 11/394,794; US 11/395,350; US 11/395,759; US 60/799,228; US 11/479,926; US

11/479,841; US 11/479,925; US 11/479,056; US 11/478,714; US 11/479,899; US 11/479,506; US 11/479,267; US

60/837,816; US 11/528,267; US 11/529,132; US 11/540,281; US 11/540,322; US 11/529,947; US 11/540,902; US

11/541,056; US 11/529,942; US 11/540,282; US 11/529,946; US 11/540,320; US 11/529,842; US 11/540,904; US

11/541,252; US 11/541,313; US 11/540,086; US 11/540,739; US 11/540,185; US 11/540,107; US 11/540,900; US

10/610,780; US 10/832,509; US 11/608,340; US 11/608,350; US 11/608,358; US 10/771,315; US 10/771,409. Other

U.S. and International Patents Pending.

VERINT, the VERINT logo, ACTIONABLE INTELLIGENCE, POWERING ACTIONABLE INTELLIGENCE, WITNESS

ACTIONABLE SOLUTIONS, STAR-GATE, RELIANT, VANTAGE, X-TRACT, NEXTIVA, ULTRA, AUDIOLOG,

WITNESS, the WITNESS logo, IMPACT 360, the IMPACT 360 logo, IMPROVE EVERYTHING, EQUALITY,

CONTACTSTORE, CLICK2STAFF, and SMARTSIGHT are trademarks or registered trademarks of Verint Systems

Inc. or its subsidiaries. Other trademarks mentioned are the property of their respective owners


3/8

Unauthorized use, duplication, or modification of this document in whole or in part without the written consent of VerintSystems Inc. is strictly prohibited.

By providing this document, Verint Systems Inc. is not making any representations regarding the correctness orcompleteness of its contents and reserves the right to alter this document at any time without notice.

Features listed in this document are subject to change. Please contact Verint for current product features and specifications.

All marks referenced herein with the or TM symbol are registered trademarks or trademarks of Verint Systems Inc. or itssubsidiaries. All rights reserved. All other marks are trademarks of their respective owners.

2009 Verint Systems Inc. All rights reserved worldwide.

Table of Contents

Verint Impact 360 Speech Analytics ....................................................................................... 1

Accuracy ................................................................................................................................... 2Efficiency .................................................................................................................................. 2Intelligence ............................................................................................................................... 4Conclusion ............................................................................................................................... 5


4/8

Page 1

Verint Impact 360 Speech AnalyticsVerints market-leading, 5

th-generation Impact 360 Speech Analytics solutions are based on a unique

technology called the Complete Semantic Index. This technology helps compensate for the limitations ofexisting off-the-shelf speech recognition by incorporating a process that leverages best of breed ASR(Automated Speech Recognition) engines with a unique categorization and indexing layer that adds

significant improvements in accuracy, efficiency and intelligence.

The first step in building a Complete Semantic Index includes phonetic processing that translates the

audio into a string of the basic sound elements of human speech (called phonemes). The output of thisinitial stage alone can provide basic audio search capability but has limited accuracy and no linguisticcontext. For this reason, Impact 360 Speech Analytics is comprised of additional processing stages that

add linguistic and acoustic mapping on the initial phonetic output, leveraging an up to 100,000 term(a.k.a. full language) dictionary, and a best of breed LVCSR engine (Large Vocabulary Continuous

Speech Recognition). Here, Verints technology, coupled with the knowledge gleaned from deploymentsacross hundreds of customer sites, comes into play. This additional stage contains rich linguistic and

acoustic models that embed extensive knowledge on how different people express themselves in contactcenter conversations across multiple verticals, both in their choice of words as well as their accents anddialects. Armed with these statistical models, the Complete Semantic Index is able to significantly improve

the accuracy compared with identifying terms based only on strings of sounds (phonemes), and provide atranscription of the call content. Verint continuously evaluates the state of the art of speech technology

and the result is a speech engine that incorporates richer linguistic and acoustic models, providing evenhigher accuracy and performance.

In the next step, transcriptions of the recently processed calls are added to an index, which is updated on

an hourly basis. This index contains all the content of all processed calls, as well as additional statisticalinformation and call metadata. This index enables the powerful and unique functionality of Impact 360Speech Analytics solutions, such as the Category Wizard, automated term suggestions and

TellMeWhy. These capabilities, compared with phonetic search solutions, are reviewed in the followingthree sections, covering Accuracy, Efficiencyand Intelligence.


5/8

Page 2

AccuracyThe unique linguistic and acoustic models that the Complete Semantic Index leverages significantly

reduces false positive results and provides improved accuracy. These layers can help distinguishbetween similar sounding words and phrases such as the word Cancel and the phrase Can Sell, which

would not likely be distinguished by a phonetics only search. Since there are only about 40 differentdistinct phonemes that make up the spoken English language there are many cases such as this one

where words and phrases may sound similar but convey very different meanings. In fact a phonetics onlysearch solution is likely to generate about 3 such false alarms (false positive term recognition) for every

word searched on each hour of audio processed. The additional linguistic and acoustic models used bythe Complete Semantic Index can reduce these false positive results by factor of 10. On the sameprobability of detection (recall), Phonetics-only search generates ten times more inaccurate results (false

positives) compared with the same search using a Complete Semantic Index. When using phonetics-onlysearch, a multiple term search or category results in 10,000 hours of contact center calls, may translate

into tens of thousands of false positives or false alarms. For example, a 500 seat contact center cangenerate on average 60,000 hours of audio per month which can equate to 700,000 calls. A single

phonetic search on this data can result in on average over 180,000 errors.

Although for the majority of terms the linguistic layers used by the Complete Semantic Index significantly

improves search accuracy (Impact 360 Speech Analytics also identifies many acronyms whileautomatically suggesting additional alternative terms), for rare non English terms and names, a phonetics

only search solution can have some advantages. As phonetic processing does not leverage a fulllanguage model it can, in some cases, identify unique names that are not part of every day contact centerlanguage. In the case where a proper name or customer

specific term is missing from the language model, Verintleverages the phonetic layer to add a phonetic match of the

name. This tuning process can help customers quickly andeasily add unique names to the model if required. As

discussed above, the search accuracy of the terms added

using this phonetic layer may not be as high as other terms,but this provides a quick and easy way to address potential

out-of-vocabulary situations. In cases where the accuracy ofthese added terms in not sufficient, Verint offers a process that

allows for the linguistic and acoustic models to be customizedto a specific environment thus bringing even non English

terms and names to the same accuracy level as all otherterms, providing higher overall recognition accuracy.

The Verint approach leverages the best of phonetics and LVCSR recognition layers providing a completesolution that produces higher overall accuracy together with the needed flexibility for contact center

environments. However, the unique Complete Semantic Index provides many additional benefitsdiscussed in the following two sections.

EfficiencyThe phonetics pre-analysis processing stage is short, which translates into relatively fast call processing.

However the resulting phonetic index presents several challenges that greatly reduce the efficiency of asolution that relies mostly on a phonetics index. A phonetic index is very similar in size to the original

audio files, making it harder to scale it to accommodate for large call volumes, and make searchingagainst it much slower (about 600 times slower than a parallel search against the Complete Semantic

Index). Each new search against a phonetic index may take anywhere between minutes to hours to return


6/8

Page 3

results on large volumes of audio, and is dependent on the number of terms searched and the audio

sample size. This makes ad hoc searching and category definition a very time consuming process. Forexample, a phonetic only ad-hoc search against a months worth of calls for a 500 seat contact center will

take approximately 70 minutes, using the fastest available phonetic technology. In comparison, theComplete Semantic Index produces new search results for any number of term or phrase combinations

within less than 3 seconds for millions of calls.

In addition to the large phonetics index, most 3rd

party solutions also create an additional copy of the

original audio, in effect tripling the total storage required compared to a fully integrated Speech Analyticssolution.

Verint has succeeded to greatly improve the efficiency of the Complete Semantic Index processingphase. Verint can now process calls in speeds of 40 RT (Real Time), transcribing 500-1000 hours of

audio per day on a single standard off-the-shelf server. In most multi-site enterprise environments, atleast one transcription server is allocated to each site, to avoid transfer of large call sets from site to site.In such environments, the total number of servers required for a phonetics only search solution is similar

to the number required for Verints Complete Semantic Index solution, which provides much higher value

and accuracy. Overall, Verint provides the most hardware efficient solution that can transcribe andanalyze the entire content of all processed calls.

The Complete Semantic Index also introduces many other unique benefits that increase efficiency. For

example, category definition is fast and easy. By analyzing the entire content of calls, the categorydefinition wizard proactively suggests correlated terms and phrases available in order to accurately

categorize calls. Often, the system will surface terms that users may not otherwise think of. Without theability of analyzing the entire content of calls users have to guess the terms they will use and do not know

how effective or relevant these terms are, and whether they may be missing significant sets of relatedcalls. For example: refund, return, cash back, my money, can all mean the same but arephonetically different, so using phonetics only search you need to know and input all terms separately.

With Verints Speech Analytics a user can input one of these terms and the system will surface thecorrelated alternatives actually being spoken in that specific environment, providing much faster and more

accurate category definitions.

Building a robust category with Impact 360TM

Speech Analytics typically takes between 30 minutes to 2hours. Defining categories in a phonetics only search solution is a long process that includes many

iterations and long waits for each new search to return results. Consequently, building a robust categorymay take days or weeks and require much more experienced users or even professional services from


7/8

Page 4

Impact 360TMSpeech

Analytics automaticallysurfaced that problemswith passwords are thekey driver of calls relatingthe self service portal

the phonetics search solution vendor. Verint also provides the ability to easily retune categories to keep

up with changes in the business environment, by automatically surfacing new correlated terms that canbe added to the category definition.

Once a category is defined, all calls are instantly re-categorized, enabling immediate drilldown into theroot causes driving this category. Verint automatically increases category accuracy by using a unique

ranking algorithm backed by self-learning capabilities. All of these capabilities are unique to VerintsComplete Semantic Index technology and are not available in any phonetic search solution that can only

return results on phonetic matches, based only on terms users know to search for.

IntelligenceLarge contact centers handle millions of calls in an ever changing dynamic environment. There are many

issues that drive calls into the contact center, and every new day introduces new reasons and challengesthat need to be recognized, evaluated and resolved. In order for a speech analytics solution to be able tocontinuously analyze this massive amount of data and surface key trends, even if they are just emerging,

it needs to provide more than just smart search capabilities. Phonetics-only search solutions cannotsurface unknown issues. Such solutions require users to know exactly what issues to search for and

exactly how they are articulated by agents and customers.

Verints new TellMeWhy functionality automatically surfaces unknown issues. For every call set,whether it is billing related calls, or calls with duration of 10 minutes an above, Verints TellMeWhy cansurface the key drivers of these calls. By leveraging the Complete Semantic Index, the system compares

the content of the specific subset of calls to the entire content of the rest of the calls, identifying terms andphrases that are unique to this call subset. TellMeWhy then clusters these terms and prioritizes the top

five root cause groups, automatically suggesting the top driving reasons for any call set. This functionalitycan help focus users attention to call drivers, whether known or unknown, that have the most impact on

their business. Answering common call center management questions such as: why did I have a callspike yesterday, what drives my longest calls, what were customers calling about yesterday, what isthe biggest complaint driver.


8/8

Page 5

Verints TellMeWhy functionality enables instant drilldown into each Root Cause call set, pulling and

prioritizing the relevant calls so users can quickly get to the information they need.

When playing a call, Verint presents a unique visual map of the call. This visual audio navigation aid

consist of an energy envelope (waveform) of the call, with the terms suggested by the system highlightedabove it to help get to the relevant segment in the call. A synchronized interactive transcription of the

entire call is also available, with terms that relate to distinct business issues highlighted in different colors.

Users can recognize additional issues in the transcription and skip directly to the relevant area in the call

by clicking on any of the terms. This visual transcript map can save over 90% of the time it takes to reviewthe call compared with a solutions that do not have a visual synchronized transcription to every processed

call. It enables a quick and efficient call review while providing easy access to additional insights andintelligence embedded in these calls.

ConclusionVerints market-leading, 5th- generation Impact 360 Speech Analytics solution leverages the full contentand context of every processed call with a phonetic layer for adding unique terms, providing the mostcomplete solution with unique advantages and capabilities, including:

1. Highest accuracy - with contextual understanding

2. High flexibility a combination of phonetic, LVCSR, and unique indexing technology

3. Instant query and categorization results no need to predefine searches in advance

4. Call relevancy ranking prioritizing the most pertinent calls

5. Proactive category wizard defining categories in less than 2 hours with minimal training

6. TellMeWhy surfacing what users may not know to search for

7. Visual map of the call quick and efficient call review with additional insights

8. Full Integrated Solution enabling lower total cost of ownership, increased security and manyWFO application synergies with quality monitoring, data analytics and reporting applications.

speech analytics - the complete semantic index

Documents