My experience with this package has been that it is a very complex software delivery in an emerging technology, and the design has not been adequately tested for user acceptance.

The package we work with has no consistency in design–the operator controls are different in different parts of the application (one designer’s work doesn;t follow the same convention as another designer for another aspect of the package.)

Also, the complexity of the application apparently escapes system designers–for our situation, the install was estimated to require 11 servers. A year and more than three times that number of servers later, we are still hoping to achieve our requirements for acceptable storage and capture.

All that said, there is great benefit to be gained from this technology. Just don’t expect the vendor to always be there to help you through the rough spots–commiting to speech analytics will require more resources than you will expect, both in infrastructure and personnel.

Point #7 is a very dangerous comment and you should really brush up on your stats knowledge before quoting specific numbers.

You completely miss the point that in order to acheive a 95% confidence limit you need to understand first the error rate and this then in turn defines the sample size required.

You can’t just quote 400 calls and state this is suitable….this is completley mis-leading.

This is a good article except that #7 is wrong and #8 over looks detection. It is important to realize that speech analytics is about accuracy & detection. Also, when determining sample size you need to account for the standard deviation.

You point out in #8 that most speech analytics vendors don’t have the accuracy they claim. If you build your categories carefully you can have “90% accuracy”…. with 15% detection. So while 90% of the results will accurately detect the word or phrase it will not detect 85% of the instances that word/phrase actually happened in your calls.

Further, the accuracy they claim is the best possible accuracy and does not tell us the variability (standard deviation) in that accuracy. For example, many systems do very poorly when dealing with accents or foreign languages.

Given the variability in detection you need a much larger sample size if you want to have a 90% confidence interval. If you’re not going to do your sampling properly why do it at all?

As you point out, it is important to have people that understand how to use Speech Analytics (and statistics). Do you want to be making organizational decisions based on bad data?

Here’s some reading on how to do sample sizes properly:
http://highered.mcgraw-hill.com/sites/dl/free/0072549076/79745/ch07.pdf