Speech Analytics is one of the fastest growing technologies in contact centres. While it may not yet be mainstream, rest assured, it will be - and, as we shall see, for good reason. That said, as with much in the field of Artificial Intelligence (AI), there is a great deal of hype that does this exciting area a huge disservice.
In this ‘The Ultimate Guide to Speech Analytics’, I shall endeavour to lift the veil on Speech and, more broadly, ‘Interaction’ Analytics! As we shall see, Analytics is already having a marked impact on two key areas - compliance and Quality Assurance (QA), and it has the potential to revolutionise other, perhaps less obvious, areas too.
In this guide, we will explore:
- What is Speech Analytics?
- What is the purpose of Analytics?
- How does Analytics work?
- Analytics use cases
- The business case for Analytics
- How to implement best practice and avoid the common pitfalls
What is Speech Analytics?
There is no one definition of Speech Analytics. However it can be defined as “software that enables organisations with contact centre operations (whether large or small, inbound and/or outbound) to have any or all of their advisor/customer calls systematically analysed”.
Believe it or not, Speech Analytics was first deployed in a select few contact centres (though they were referred to as call centres back then!) around the turn of the millennium. It was the advent of speech recognition technology which, perhaps even more incredibly, emerged in the early 1950s, that paved the way for Speech Analytics - without the means to codify spoken words and phrases, Speech Analytics would, quite simply, be impossible. It is also worth noting that Speech Analytics goes beyond recognition and analysis of the spoken word. Speech Analytics also tracks factors such as tone, sentiment, rate of delivery, silence and people talking over each other.
Speech Analytics is itself a subset of what is variously termed ‘Interaction’ or ‘Engagement Analytics’. Just as we witnessed the evolution of the omnichannel contact centre from its more humble origins as the call centre, Analytics has also been developing to reflect these changes. Thus, many (but by no means all) solutions now permit analysis of other non-voice contact channels, e.g. email, Webchat, Social Media, etc...
What is the purpose of Speech Analytics?
Most would agree that accurate and timely data analysis is a key contributor to organisational success. It is vital in order for informed decision making to take place. However, historically, there was a problem...
Data falls into two broad categories - ‘structured’ and ‘unstructured’.
|Structured data||Unstructured data|
Typically anything that has a discernible and quantifiable pattern or quality that can be recorded meaningfully and precisely as a value, e.g. a number or date.
The majority, at up to 80% - is everything else!
Freeform speech is an excellent example of unstructured data. A conversation may be fluid and dynamic but, as long as the listener has the knowledge and skills to interpret it, some level of understanding and, hence, insight should be possible. The purpose of Speech Analytics is to provide an artificial means to automate much of this process reliably and at scale. It enables the user to convert unstructured data held within calls into a usable format that may itself be augmented with structured data (e.g. sales outcome, Customer Satisfaction score, etc.) to enable in-depth analysis and yield insight.
As we shall explore later in use cases, Speech Analytics may be deployed to address a number of requirements. In Contact Centres it is more commonly used to track and flag compliance breaches and for some, underpin quality assurance processes.
How does Speech Analytics work? And why you need to be careful.
As we have seen the purpose of Speech and Text Analytics is to take unstructured interaction data and to render this into something more usable that may then be manipulated to yield insight.
There are three main timeframes in which Engagement Analytics is applied, these are:
At present, this is the most common form of Analytics. Calls and/or text interactions are made available to the Analytics platform either by secure on-going or batch-transfer. Dependent upon a number of criteria (e.g. processor specifications, Analytics solution, etc.) reporting may be as fast as within a few hours or, more commonly, by the next day.
Sometimes there may be a requirement to process interactions that took place days, months or even years earlier. An example of such a use case is the redaction of historic ‘toxic’ PCI calls.
3. Real-time Speech Analytics
A growing trend within the Interaction Analytics arena is Real-time Speech Analytics, however at present, it is undoubtedly one of the most over-hyped applications of Analytics for a variety of issues.
However there's currently a number of challenges associated with implementing speech analytics, especially when applied in real-time.
Issues with with Real-time Analytics
Over-hyped by many software vendors, there are five main issues with Real-time Analytics:
- Barriers to entry. The complexity (and cost) of integrating Speech Analytics with live channels.
Accuracy. The accuracy of Speech Analytics at the individual interaction level may be questionable and, for example, for compliance purposes, there is a high risk of false-positives and negatives.
Manual effort. At present, alerts and prompts are likely to require manual creation in the first place. This can be time-consuming and costly, and prone to error, incompleteness and over-complexity. Furthermore, such alerts and prompts need to be proactively managed if processes, rules or best practice changes.
Agent Engagement. Having pop-ups appear can be very distracting for an agent and damage the natural flow of their call.
Latency - how real-time is real-time? If alerts or prompts are not provided in a timely manner, they can be worse than useless. Leading delays, poor service, stress and, where applicable, leading to potentially catastrophic failures, e.g. non-compliant interactions.
Besides these technical challenges, analytics, notifications and alerts don’t change front-line behaviour by themselves; there is a danger of using ‘real-time’ as a sticking plaster, rather than seeking to address the fundamental underlying issues. We all should know that it is employee engagement, coaching and development that drive real improvement. For the majority, it may be wiser, and considerably less expensive, to invest in better training, coaching, alternative technology and processes.
Expanding on challenge number two, the following is a genuine example of a transcript (and it is by no means atypical).
When people are first exposed to transcripts, their usual immediate response is, “they’re awful!” For certain use cases, they would be right... Where accuracy is key, e.g. for real-time compliance alerts, words and phrases may be missed altogether or falsely reported.
However, where Speech Analytics is being applied to help flag potential areas of concern for a human evaluator to assess, such imperfections and challenges can be mitigated. It is often the case that just a few tell-tale words, phrases or data-points are all that is required to indicate that there may be a problem. A good example is in the assessment as to whether agents are consistently asking for appropriate information in the Identification and Verification process. The words and phrases associated with such a measure are usually readily configured and spotted even if the transcript isn’t anywhere near 100% accurate.
How do customer service and sales team use analytics today?
While many challenges remain, the Speech Analytics market has been growing exponentially in recent times. In the early years, Analytics was deployed predominantly by large-scale operations in highly regulated industries such as financial services (banking, insurance and debt-collection businesses remain major users). Utilities and telecommunications companies soon followed. However, with greater accuracy and reducing costs, Speech Analytics is now found in most areas - from travel to healthcare, from charities to public sector. Furthermore, people are finding ever more uses for it.
Within today's contact centres, speech analytics is commonly used to:
- Underpin compliance and script adherence
- Partially automate elements of the quality monitoring process
While many may never take it further than the above, this only represents a fraction of Analytics’ true potential as we are about to explore.
At this point, it is worth noting that Analytics should only ever be viewed as a tool to empower Quality Assurance teams, not replace them. To support the heavy lifting and draw their attention to potential issues and areas of interest which they may then investigate.
In turn, this should enable them to act on the generated insight; driving positive change in agent behaviour, updating e-learning material and helping to support coaching sessions, for example.
In addition to this, speech analytics can be used to:
1. Out-of-the-box capabilities
Out-of-the-box, Speech Analytics can be used to identify excessive silence and where agent and customer are speaking over each other. Whether across the operation or at an individual level, these are tell-tale signs that there are problems. Silence is often evidence of training, process or system issues, while talk-over is frequently an indicator of frustration, pressure, anger and a poor customer experience. Identifying and addressing the causes of these can dramatically improve: Average handle time and operational efficiency (shorter duration, better calls) Customer experience and satisfaction (CSAT & NPS) Outcomes (reduced churn, increased sales, propensity to recommend)
2. Compliance & script adherence
Whenever it is essential to ensure that contact centre agents are adhering to a script, Speech Analytics provides the ideal solution; regulated environments represent a prime example. Analytics may be deployed to detect and flag variances from required scripting that may indicate non-compliance or worse, misselling. However, the caveat regarding accuracy, false-positives and negatives should be borne in mind - Analytics is not infallible and should only ever be used in conjunction with human evaluators.
3. Partial automation of the Quality Monitoring (QM) process
Evidence suggests that most organisations are still only assessing 1-2% of advisors’ calls each month; figures tend to be even worse for text-based interactions, e.g. Webchats, email and Social Media. Few would disagree that conducting regular, objective assessments of the quality of interactions between advisors and customers is a good idea. However, generating a sufficient sample size to permit robust analysis of an agent’s performance can be a challenge for many.
This is where Analytics comes into its own as it may be used to do the preliminary ‘grunt work’.
However, it can only take you so far - save for the ‘Hawthorne effect’. Analytics doesn’t change agent behaviour or deliver performance improvement by itself.
An analogy would be to liken Analytics to early-warning radar. Radar, on its own, is just the starting point; should it detect an anomaly you then need to scramble your available defences to allow you to investigate and deal with the situation appropriately. Returning to the QM process, this represents the point at which it is often most appropriate to deploy human evaluators. Thus, Analytics may be used to automate the initial process in a statistically robust fashion and thereby focus the QM team’s resources in a far more targeted and rigorous manner.
It should also be noted that, Speech Analytics is incapable of conducting the more nuanced analysis of a skilled QM evaluator.
Analytics is good at assessing binary measures - e.g. did an agent use an appropriate word or phrase at a given point of a call: ‘yes’ or ‘no’ (or ‘Not Applicable’)? However, it is not so good at, or, indeed, may be incapable of, scoring on a sliding scale where a true understanding of context is required.
4. Identification of ‘what good looks like’
As we have seen, Analytics may be used to underpin efficiency, customer service and, hence, satisfaction. This is certainly the case when deploying it in order to identify ‘what good looks like’. However, in sales organisations, it also has the potential to help drive new sales and customer retention. For example, analytics can uncover what the best and worst performing agents are doing differently during calls. In so doing, it is possible, for example, to devise coaching and training programmes to seek to replicate actions of the best-performing agents in order to deliver improvements.
5. A/B Testing
A/B testing is a well-known methodology whereby two approaches may be trialled side-by-side to determine which variant performs better. In the contact centre, this can be particularly valuable for sales and service approaches. In this context, A/B testing may be underpinned with Analytics in order not only to identify the best approach but also to inform precisely what it is that is delivering, for example, an improved call conversion rate.
While there are many advantages and use cases for speech analytics, the penetration of this technology is still relatively small with many unable to build a strong enough business case or find a suitable vendor.
The benefits and costs of speech analytics software
Clearly, any business case has to be built according to the relevant prevailing factors and considerations, so it is hard to be prescriptive here. However, many cite six benefits for deploying Analytics:
- Risk mitigation and compliance
- Quality improvement
- Conversion rate improvement and thus customer acquisition
- A better understanding of the customer via an increase in searchable data
- Customer retention and a reduction in churn
- Additional resource by creating more time for human evaluators to drive performance improvement
For many years, advanced Analytics vendors only really targeted organisations with larger scale contact centre operations, often citing 250 seats as being the minimum viable scale in most circumstances. However, in recent years, not least with the increasing move away from premise-based to Software as a Service (SaaS), ‘cloud’, deployments, a range of pricing models are commonly available.
|Pricing model||How it works|
Per ‘processed hour’
Based on the maximum number of hours of call recordings the platform guarantees to be able to process on a daily basis
Per agent, per month
The number of agents who's calls are being analysed. (Note: this often comes with a max cap)
Fixed term / Perpetual
|Traditionally a one-off payment that provide a licence for a fixed-term or for the lifetime of the version. (Note: This is increasingly uncommon)|
In addition to the cost associated with the licence itself, some speech analytics vendors may also charge for:
- Call extraction
- Integrations with existing systems
- On-going support charges to get you up and running and to maintain the software
- Additional contact channels for analysis
- Professional services such as an Analyst to help make sense of the data and identify areas of improvement
How to avoid the common pitfalls when deploying any speech analytics software
As highlighted, Speech Analytics software is a tool rather than the complete solution to a problem. Frankly, analytics is only as good (or bad) as the people programming, interpreting and acting upon its output.
If an organisation is to reap the full benefits of deploying Analytics, the following points flag some key considerations regarding best practice and potential pitfalls:
Analytics is sophisticated software
In order to use it, you will need, or have ready access to, an analyst or analysts. This may seem obvious but, all too often, deployments fail as the result of inappropriate individuals being assigned to manage solutions. Logical and inquisitive thinkers with a good grasp of complex spreadsheets and a good understanding of basic statistics, often make good analysts...
Good software and analysts aren’t enough
Organisations will also have to have the processes, people and, where applicable, technology in place to act on insight but also to determine what to investigate in the first place. In short, appropriate stakeholders must be engaged from the outset and on an on-going basis if deployments are to be successful...
Good, usable data is needed
Analytics’ success is predicated on the availability of good, usable data, as well as the platform’s ability to process this data in a timely fashion. For example, for Speech Analytics, this means having access to good quality, unencrypted audio with the necessary associated metadata, as well as the platform having the appropriate ‘language packs’ and, of course, the solution will need to be (re-)configured and audited, as required, and on an on-going basis, e.g. confidence levels will require management and ‘fine-tuning’ must be maintained.
Don’t attempt to overcomplicate things
Incremental deployment is far more likely to yield the best results. It is often better to start with one relatively simple area (e.g. identifying silence and talk-over) that will deliver a quick win, even if it is not necessarily the one that will, ultimately, result in the greatest gains. This helps to familiarise the team with the platform and to gain stakeholder buy-in at an early stage.
Don’t fall into the trap of ‘throwing the baby out with the bathwater’!
A classic example of this is the mistaken belief that Analytics can replace, wholesale, existing QA processes and solutions; it can’t:
Yes, Analytics can be used to assess and report on up to 100% of customer interactions, thereby rendering the process statistically sound and removing the onerous, labour-intensive and costly first step human evaluation. However, as stated previously, this should be seen as a means of freeing-up valuable resource to investigate flagged anomalies, and, as appropriate, drill-down into ‘root-cause’.
Not forgetting, that the organisation must then have the wherewithal to act on such insight, e.g. coaching, training, process improvement, etc..
Put simply, and despite what many other software vendors state - Analytics should be only ever be used to augment human capabilities - recognising both have strengths and weaknesses.