How Accurate Are Your Speech Analytics?

22nd June 2021

Speech analytics is one of the most powerful tools in unlocking value in your business. However, having said that, how do you know that your speech analytics tool is delivering accurate results? While no system can be perfect, is there an acceptable level of error and how do you track and fix these errors to ensure that your strategic decision making is based on the right results?

A case for error tracking

The strength of a speech analytics tool is in being able to deliver accurate results based on specific search criteria. In technical terms, this is called a search pack and unless it is customised to a clients’ specific requirements it’s unlikely to generate the desired results. Additionally, because errors will occur, it’s essential to control for these errors and ensure that they are within an acceptable range.

As an example, on a particular project, a speech analytics search was run with two different search packs to determine which would be more accurate. The search was related to sales pitches within call volumes. The first pack returned a result of 1900 of 30000 calls. However, the margin for error was such that the result could actually vary from as low as 1600 to as high as 5500 calls. That is a very high range when you are looking for accurate data on which to base a strategic decision. The second search pack delivered a result of 2750 with an error range between 2600 and 3300. While there was still an error in the second result, the range of this error was significantly smaller. If you were making strategic decisions based on this data it’s almost certain you’d opt for the second pack with the smaller error range.

How to reduce your error range

While many speech analytics tools provide a variety of features, the focus is primarily on call fidelity and audio quality as well as transcription accuracy. However, when a system is installed it’s rarely optimised for a specific business, meaning that the search packs run are often generic and more likely to generate a high error range.

For more accurate results it’s essential to utilise technical expertise that has a level of understanding of your business and how search packs are generated. They will be able to use robust methodologies that can track at least 8 different common error types. Once the errors have been identified they can be fixed and searches can be customised so that they reflect more accurate results.

A key point is to have a thorough understanding of both your business and your contact centre. What phrases are commonly used by both advisors and customers? For example: If you want speech analytics to identify how often sales pitches are introduced into customer calls the system needs to be able to listen out for key phrases.

Chances are the advisors will not be asking the question as directly as “So do you want to buy that?” They are more likely to use phrases such as: “Would you be happy with that as a solution?” “Would you like me to go ahead and secure that for you?” “What many customers have found useful is this product ….” “Would you like the 12 or 24-month guarantee package with that?” To increase the accuracy of search results it’s also important to categorise possible customer responses. A few examples of this could be: “I’ll go with the 12-month option.” Or even a negative response such as “No thanks” could be an indication that the contact centre advisor made the effort to make a sales pitch.

Another way to reduce the error range is to combine searches with metadata. This is the non-verbal information in a call. A long silence, on-hold time, frequent pauses or hesitation in the conversation with the contact centre advisor, often tells you as much as the conversation itself. For example, if you’re looking to reduce average handling time, then doing a search for calls that last more than 150% of the AHT could reveal problems in the system or with the knowledge and skill levels of advisors that might need to be addressed.

Get it right from the start

Because speech analytics is something that generates data over time it is useful for historical analysis. However, for this to be effective you need to ensure that the right search packs are used from the start, otherwise, your baseline will be flawed. Remember that customization is key to getting the right search results for your business with the smallest error range. It’s worth investing in the technical expertise to get this right because without it there is no way of knowing just how accurate the data is and if you are making strategic decisions for your business based on fiction or fact.

Author: Lee Mostari, Director of Insights and Analytics, Davies Consulting

First published 15th June 2018


    Keep up to date with Davies