What Does Speech Analytics Software Actually Do?
Speech Analytics software comes in a diverse range of flavours, all working “big data” style, to uncover insights previously hidden in call recordings. Many enterprises and smaller contact centres are finding niche uses for this technology as well as broader strokes like trend analysis, compliance and sentiment detection. All of which is very interesting, however we’re here to answer what is happening under the hood? What does Speech Analytics software do exactly?
Well, there are two ways speech analytics software breaks down the data. Speech Analytics technology will use one or both of the following methods…
1. Phonetic Indexing
The first type we will broach is called Phonetic Indexing. The process here is simple, the speech/audio is broken down into strings of phonemes, the basic units of speech. Because phonemes simply consist of uttered sounds, the indexing is unaffected by background noise, languages, dialects or speaking styles. Phonetic indexing creates the truest representation of spoken audio and enables the fastest, most accurate access to the data contained in audio recordings. The problem with phonetic indexing is that you need to know what you’re looking for, and it can take a considerable amount of time to develop a speech library that will get the results you need.
2. LVCSR (large-vocabulary continuous speech recognition)
Speech to text is the other big player and while it’s not as fast or as accurate, it does have some advantages. In the industry we call ‘speech to text’ large-vocabulary continuous speech recognition or LVCSR. LVCSR recognizes uttered sounds much like phonetic indexing, but subsequently, it matches combinations of phonemes against linguistic models containing a large human-language vocabulary to build a complete database.
Most LVCSR engines provide a transcript that gets instantly parsed and sorted into snippets. This allows analysts to start turning unstructured data into something meaningful. This method has an aptitude for surfacing new business issues and analysing trends in near real time.
Speech Analytics Accuracy
It’s extremely difficult to find exact numbers on the accuracy of speech analytic engines. However phonetic indexing has an advantage over LVCSR in that it’s easy to adjust for different dialects or specific product names as you’re dealing with each phoneme. As you develop your speech analytics library you will account for this, balancing detection and accuracy. It is a simple recognised axiom in speech analytics, as accuracy goes up, the detection rate goes down. This is because you’re omitting results. If you set a broad net you’re likely to include phrase matches that are not exactly what you’re looking for, however setting the accuracy too narrow and you’ll miss important data.
Depending on a few factors like quality of the audio and the phrase you’re searching for, an accuracy limit between 40% to 70% will yield good results. However, depending on the size of your libraries, you may be tweaking and optimising for weeks or months before you’re confident in the results.
The other factor that determines accuracy is how long the actual phrases that you’re searching for are. Phrases with a long tail are going to be matched much more accurately regardless of the speech analytics software you choose.