Feedback from the Speech analytics webinar

In our Speech Anlaytics webinar we promised to publish all of the feedback.  The event was a great success.   Over 232 people registered for the webinar, 111 people dialled in on the day and a further 50 people have watched the recorded Webinar.    Here are a list of the questions and answers.  Questions and answers

Q1.  Please cover off the reliability of the word recognition. Currently not getting firm feedback that it is accurate, particularly across a highly diverse business where you would need to socialise a large number of different words or phrases.

A1.  We weren’t able to cover this off in much detail in the Webinar, but the key point for Nexidia is that we use a phonetic indexing technology to structure and then search/query the calls. We explicitly do not do recognition, the system is only aware of the existence of the basic sounds of English, not the words. This means that we do not need to define, create and modify a dictionary of the words we recognise.

Again, when we want to search for the existence of a word or phrase, let’s say “Fourteen-day cooling-off period”, the system will decompose this phrase in to its underlying sounds, phonemes, and then search the index for this pattern.

The 2 main advantages of this approach are speed; we can index up to 20,000 hours of audio in a day, and dictionary independence, you can theoretically search for any word, including names, slang, abbreviations and shortenings and without the time overhead of dictionary creation, you can do this on day one.

So far we have run projects in the UK with outsourcers, retail banks, insurers, retailers, regulators, government service providers, utilities, mobile phone companies and not run into any issues with regards to reliability or accuracy.

Q2.  Up to how many files can be captured?

A1. There is no real limit as to the volume of audio that can be captured and processed. We have people indexing hundreds and thousands of hours per day. We have tested a system to process 20,000 hours of audio in a single day. If you have an average talk time of 410 seconds, this would mean that you could process in excess of 175,000 calls per day.

Q3.  What is the success rate on detecting key words?

A3.  I would love to be able to give you a simple answer, but in truth in depends on a number of variables that includes what you are searching for, i.e. whether it is a simple word or a phrase, and the quality of the audio. Typically we do not find ourselves looking for simple words within call centres (even though we have in litigation support engagements). Within call centres we end up looking for a phrase, often with some associated contextual logic, that will highlight an issue, behaviour, process or whatever, that we are investigating. The aim is to provide data for 2 reasons, firstly a set of calls to listen to as evidence and then secondly a larger set of calls to quantify the size/scale of the issue.

Q4.  Does the system analyse the call audio for the key phrase billing question, or is it part of a process for all calls from the billing team?

A4. At one level the system can search all calls for the phrase “billing question”. Alternatively, if the call recording has associated meta-data that can show which team/queue/agents handled the call, we can also use that as a parameter to narrow the search.
Again as we perform root-cause analysis, we use both the call audio data and related meta-data to help drill down to the area we are interested in.

Q5.  Would the words ‘billing question’ actually have to be said, or could you have a broader view e.g. billing query, question, issue, problem?

A5. I could perform a one-off search for “billing question” and then I would look just for those exact words. Alternatively I could define a repeatable billing question search that was essentially an OR statement looking for “billing question”, “question about my bill”, “this showed up on my bill”. “unsure why this is on my bill” and so on.

Q6.  How accurate is the search – will the system only identify that word or will it bring up words that sound similar to it?

A6.  In answer to your question regarding similar words, there is a specific feature of what we do that helps on this. As we never create a text transcript of the call, we never make a hard binding between a set of sounds and a specific word. This means that you can run searches with differing confidence levels depending on the business need. If you need to get a sample of calls to play to a process owner, then your tolerance to a hit is different to someone who is listening to audio from a regulatory compliance point of view. By allowing a user to vary the confidence level, we can search for an exact phrase by setting a high confidence level, or we can find a phrase, plus calls that are similar if we set a low confidence level.

Q7.  Most of our calls are outbound generating appointments for our clients – has the system been used in this environment?

A7.  Yes it has, and it works equally well in this environment. We installed in a collections organisation late last year and they are using it to improve the Promise to Pay and actual Payment rates.

Q8.  Does this application have a minimum size target market, i.e. must have more than X number of agents?

A8.  There is no lower limit, but in practical terms it is at around the 90–120 seat range. That said, it also depends on the value of the transactions to either the supplier or their customers.

Q9.  Would this replace our current call recording solution or work alongside it?

A9.  The system does not do any call recording, it extracts the call from the call recorder, creates the index files and then performs the analysis.

Q10.  Can Nexidia be used on calls recorded from any recording system? What format does it support?

Yes it can. We have interfaces to most leading call recorders including Nice, Verint, Witness, Eyretel, Red Box Recorders, eTalk, TISL, Dictaphone, Mercom. Audio format is generally any commonly available audio format including, but not limited to: .aif,  .aifc,  .aiff,  .asf,  .au,  .aud,  .avi,  .flv,  .mov,  .mp2,  .mp3,  .mp4,  .mpeg,  .mpg,  .mpg4,  .nmf,  .qt, .vox,  .wav,  .wma & .wmv

Q11.  Do you think the high targets of low talk time has an impact on giving quality during a call? In a bid to meet targets, calls are rushed and quality is compromised. How can one strike a balance?

A12.  This is a real problem, partly driven by averages. If you cannot categorise your calls, then your only AHT measurement is the overall average for all calls. In practice this is often a meaningless number. It is better to understand why customers are calling and then make changes.

You may decide that you want to present your customers with a self-service option for simple processes. That way, you can start to improve the efficiency of your operation, whilst allowing call handlers the time to spend on the calls you want them to handle. These may be more complex calls, higher value calls or whatever. Once you really understand what is driving your customers to call, then you can really understand where there is an opportunity to lower handle times without impacting quality.

Q13.  If you have a contact centre with several departments, can you set separate measurements for each individual department?

A13.  Yes, we have the ability to filter calls based on the call meta-data so you can filter down to departments, teams, sites.
Each department can set up unique views that only take their calls and the present data on what is important to them. So a sales team would structure their reports differently to a customer service team.

Q14.  How does it deal with background noise, does it interpret that as talk time in the reports?

A14.  No it doesn’t. In fact in most calls, the non-talk periods often have high levels of background noise. The system filters these out.

Q15.  Can it be used for inbound and outbound calls?

A15.  Yes, inbound, outbound, transferred calls it doesn’t matter. The system is essentially audio agnostic. All it needs is a audio file, say a .wav file. It will then process it to create an index file and run defined searches against it.
You can then additionally use direction, or transfer or whatever as an additional reporting filter.

Q16.  Can speech analytics be used in a real-time environment e.g. trigger a prompt to an agent?

A16  This is dependent on the underlying speech engine. I cannot answer for other vendors, but Nexidia’s core phonetic engine can and does work in a real-time environment. We are currently at the early stage of developing and deploying this capability into the call centres.  In fact the core engine which originated from a research project in Georgia Tech in Atlanta in 1996 has always been able to process in real time.

Q17.  Does this software interact with screen recording capture as well as voice?

A17.  The system performs all of its analytics on audio files, so no, we do not use the screen recordings. That said, if screen recordings are present, then we replay them as part of the Forensic Search Client, so if you had a query trying to identify a specific process, say the creation of a direct debit, then we would be able to take you to the point of the call where this occurred and show the associated screens.

Q18.  How do you gain insight without the cost of listening to calls. Surely you still have to listen to the calls to pinpoint the right calls to listen to?

A18.  You do not completely get away from listening to calls; in fact we believe that it is critical to listen to calls. What we do is use our technology to enable our customers to perform rapid and accurate root-cause analysis to really understand and quantify what is actually happening in their call centre applications.

Q19.  How is the system costed; number of phone lines, etc. and how long does it take to implement?

A19.  The system pricing is based on the volume of audio processed per day. A 250-hour-per-day licence would allow you to analyse 2,500 360 second calls per day.

Our engagement methodology means that we recommend that customers initially start with a 120-day QuickStart engagement. This is based on an on-demand hosted service and we typically deliver results in 2–3 weeks. The aim of the project is to deliver business value through the project itself and also to develop the business case for the ongoing deployment of a speech analytics system.

Q20.  What percentage of all calls need to be recorded to achieve the success that Ventura witnessed?

A21. I believe that Ventura were analysing about 15%–20% of the calls for the area under analysis.

Q22.  Can you pre-specify which calls you want to use the analytics for, e.g. choices made in an IVR menu?

A22.  Yes, provided you can associate the IVR selection with the call recording.

Q23.  With the Nexidia product, where does the organisational structure come from? Can the system integrate with an internal system or does the Nexidia product duplicate employee administration?

A23.  At the moment the system does automatically draw in data from clients’ internal systems. However, we can develop specific integrations if required. We have a customer whose call recorder stores agent ID, but not the team leader. When we import the call data, we perform a separate loo- up to populate the team leader data.

Q24.  How do you connect/integrate the speech analytics?

A24.  The integration is as follows. We take a copy of the audio file to create the index and we typically interrogate the recorder database to retrieve the call meta-data, i.e. agent ID, CLI< DNIS, Direction, Date, Time, etc.

If you would like to watch a replay of the event, click here to watch the recorded Webinar

If you have any more questions and answers you can ask them at our LinkedIn group.