WIXLIB

Deciphering Voice of Customerthrough Speech AnalyticsDiscover Customer Insightsand Improve CXTHOUGHT LEADERSHIPW H I T E P A P E R90% of all customerconversations are happeningon the phone. That’s 56 millionhours of customer phone callsevery day!Source: Gartner

Contents1.Introduction .22.What is Speech Analytics? .22.1.Challenges analyzing speech data .22.2. Algorithms for transforming speech to structured data .32.3. Key components of a speech recognition system for Analytics .32.4. The Technology comparison .53. Organizational Use Cases For Speech Analytics .64.Solution Landscape: Vendors, Product and their market share .65.Recommended Architecture for Speech Analytics .76.Case Study: Predicting NPS in Health Insurance .87.Conclusion .8

1. Introduction For most retail businesses, customer interactionvia call centers is a very significantcommunication channel. Organizations typicallyreceive thousands of customer calls every day.According to an industry report, over 56 millionhours of conversations (nearly 420 billion words)are spoken a day in call centers worldwide. If theaudio data thus collected can be aggregated2. What is SpeechAnalytics?Speech analytics is a powerful tool for analyzingrecorded calls, structuring customer interactionsand gaining insight in the hidden information. It canbe used for audio mining, speech categorization,intelligence extraction, decision making, monitoringagent performance.and analyzed, it can yield quality insights intoIf applied correctly and used effectively speechcustomer expectations, preferences, serviceanalytics can help improve service quality, reduceissues & product usage. While speech analytics isoperating expenses, boost revenue, and reducenot a new technology to the market, most of thecustomer attrition. If integrated well with overallbusiness executives are still skeptical about thestrategy it can help businesses drive product &value it can add.process innovation leading to significant market This whitepaper aims to illustrate basictechnologies used in speech analytics, their usecases and how ROI from speech analyticssoftware can be maximized.differentiation. However there are significantchallenges in transforming speech data to astructured form which can be subjected to furtheranalysis.2.1. Challenges analyzing speechdata The speakers differ in speaking style and speed,gender, age, dialects, physical attributes (such asvocal tract). Any speech recognition system has totake all these features into consideration. Forexample “service provider” may be recognized as“serve his provide her”.Unstructured Humans in addition to speech, also communicatevia facial expressions, emotions, postures and eyemovements, and these are missed by anautomatic recognition system (ASR). While interacting in the real time environmenthumans encounter lot of unwanted sounds callednoise and these need to be filtered from thespeech signals. Homophones (i.e. words that are pronounced thesame but differ in meaning. For e.g. two & too) andword boundary ambiguities pose a major problemto speech recognition systems.analytics.rsystems.com 2

The acoustic waves change with the properties ofthe channels used.process for conversion of unstructured voice data The continuity in speech leads to problems relatedto identification of word boundaries.from written languages.To fully leverage the information encapsulated incustomer calls, you need to transform the interactiondata in the audio files to a more structured datawhich can be queried by analysts & can be consumedby sophisticated machine learning algorithms.direct phrase extraction achieve this by deploying asome or all of the following components: anacoustic model, grammars, a language model andrecognition algorithms. These components arestandard in any speech analytics software today.But understanding them is important as thequality of speech analytics is dependent on howthese components are configured and whatalgorithms they use.Phonetic IndexingRoot Cause AnalysisSpeech-to-Text(LVCSR)Emotion DetectionDirect PhraseExtractionScript AdherenceTalk AnalysisSpeech Analytics AlgorithmsValuable Insights forcontinuous improvementStructured OutputAudio and TextInteractionsphonetic indexing, speech-to-Text (LVCSR) andspeech recognition system which comprises of2.2. Algorithms for TransformingSpeech to Structured DataTransfer me tothe head officeto a structured and more useful information.The different speech analytics approaches such as Grammatically spoken language is very differentUnstructured DataFigure 1 gives the complete high-level flow of theFigure 1: Flow for transformation of unstructured to structured data in speech analytics2.3. Key Components of a SpeechRecognition system for Analyticscomponents of a speech recognition system areshown in Figure 2.Acoustic model: an acoustic model represents therelationship between an audio signal and thephonemes or other linguistic units that make upspeech. An acoustic model contains statisticalrepresentations of each of the distinct sounds that72% companies believe that speech analyticscan lead to improved customer experience,makes up a word. An acoustic model is created using68% regard it as a cost saving mechanism, 52%a speech corpus and training algorithms that createrespondents trust that speech analyticsthe statistical representations called Hidden Markovdeployment can lead to revenue enhancement.Models (HMMs), for each phoneme in a language.Each phoneme has its own HMM, and HMM is oneSource: Opus Research Surveymost common type of acoustic models. The differentanalytics.rsystems.com 3

househouseHOUSAND [HOUSAND] hh aw s ax n dHOUSDEN [HOUSDEN] hh aw s d ax nHOUSE [HOUSE] hh aw sHOUSE'S [HOUSE'S] hh aw s ix zAcoustic waveformHMM ModelGrammar&LMsPronunciation dictionaryAcoustic ModelFigure 2: Different components of a Speech Recognition SystemLanguage-specific acoustic models are usedthe number of possible words that need to bedirectly in the Phonetics indexing approach beingconsidered at any one point in the search. This resultsused for speech analytics. The basic recognition unitin faster execution and higher accuracy. Tri-grams arefor this approach is a phoneme and it is a dictionarythe most commonly used LMs in ASR. Theindependent approach. It allows users to queryprobabilities of n-grams help in determining whichphonetic strings and perform orthographic searchn-gram is more probable in comparison to the otherusing a pronunciation model.similar n-grams. For example the P (I saw a van) P(eyes awe of an) for correct recognition of the phrase.Language model: a statistical language model(SLM) consists of a list of words with their probabilityof occurrence. It is used to restrict search by limitingThis is calculated using a language model similar towhat is shown below.P(I s ) 0.67P(eyes s ) 0.25P(saw I) 0.63P(a saw) 0.5P(of awe) 0.45P( /s van) 0.6P( /s an) 0.15P(awe eyes) 0.33The large-vocabulary continuous speechwhereas, the ‘vocabulary’ file defines wordrecognition (LVCSR, also known as speech-to-textcandidates in each category, with its pronunciation(STT) or word spotting) speech analytics approachinformation. For illustration, consider the sentenceuses both language-specific acoustic models and“I’ll take one house please”. The ‘house.grammar’ inlanguage models. The basic unit for this approachthe BNF would be as shown below. ‘S’ indicates theis a set of words, and these are generally bi-gramssentence start symbol. The rewrite rules are definedor tri-grams. In order to map words to phoneticusing ‘:’ symbol.forms it also uses a pronunciation model or adictionary. This is a dictionary dependentS : NS B HMM SENT NS Eapproach.S : NS B SENT NS ESENT: TAKE V HOUSE PLEASERecognition algorithms: perform speechSENT: TAKE V HOUSErecognition based on written grammar. ThisSENT: HOUSE PLEASEgrammar describes the possible patterns of words.SENT: HOUSEThe recognition grammar is generally given usingFRUIT: NUM HOUSE Ntwo files the ‘grammar’ and the ‘vocabulary’ file.FRUIT: HOUSE N 1The ‘grammar’ file defines category-level syntax,analytics.rsystems.com 4

The vocabulary file contains definition of each word defined in the ‘grammar’ file. The partial ‘house.voca’ filefor the ‘house.grammar’ file is given below.% NS B s sil /s % NS Esil% TAKE VI'll takeay l t ey k% PLEASEpleasep l iy z% HOUSE Nhousehh aw s%HOUSE N 1Houses% NUMonetwow ah nt uwhh aw si z2.4. The Technology ComparisonFILLERFILLER% HMMfmw eh l Audio Processing: The audio needs to bereprocessed for new words added to theThere has been a long-standing debate on the meritsdictionary. This proves to be a time consumingof these approaches. Several factors are consideredand a costly process, as experts are also neededwhile choosing the technology. There is no one bestfor entering words in the dictionary. The audio isapproach that may be pointed out, however, theprocessed just once in the dictionaryapproach may be chosen based on the requirementsindependent approaches.as discussed below. Speech transcripts: The dictionary independent Targeted listening or calls of interest: If you aim toapproaches (phonetic), due to absence of alisten to calls containing specific keywords, then alanguage model cannot be used to generate asystem with dictionary dependent (LVCSR) may bemeaningful orthographic transcript of speech.preferable. The dictionary dependent approachescan recognize words that are already in the lexicon. Out of Vocabulary (OOV) words: If it is very likely Precision vs. Recall: A higher precision rate maybe obtained with a dictionary dependentsystem on words that are already in thethat you encounter new words in your searchdictionary, but the recall rate suffers as there aredomain, then dictionary independent approachalways some missing words. This also results inshould be preferred. OOV handling is a major issuehigher error rates. Phonetic approaches havein the dictionary dependent approaches.low precision but higher recall.FeaturesSpeedRecallPrecisionError LowHighHighLowApproachTable 1: Comparison of the approachesanalytics.rsystems.com 5