Dialogic Blog

Which speech analytic engine is right for your application?

by Vince Puglia

Sep 12, 2017 4:18:23 PM

Confession - last week while entering my dark hotel room, I found myself uttering ‘Alexa, lights on’ – to my disappointment the room remained dark....


Indeed real-time speech analytics and natural language processing are changing human behavior (at least, it’s changing my behavior) and we seem to be at the forefront of this paradigm shift, but with so many options, which speech analytic engine best? A simple search would generate an abundance of varying opinionated blogs, how-to's and even some voice assistant battle videos (one of my favorites) but still no definitive unified answer.

Recently the Dialogic applications team looked to leverage real-time speech analytics and natural language processing with our video conferencing solution to create a ‘conferencing valet’. The idea was to integrate the speech analytics service as a passive listening participant and trigger actions based on what it heard – in our case, it would trigger visual advertisements in the chat window. We needed a cloud service that could quickly and accurately translate the speech of the conference attendees into text then be able to extract specific intents from the speech for actions. This led us to evaluating several vendor offerings and while in the end we decided to integrate using IBM Watson - the short and anti-climactic answer to which real-time speech analytic engine is the best is…… depends. 

Now let me explain before you close out this blog – the reason for the non-decisive answer is because each vendor has both strengths and weaknesses, which should be considered, based on the application use case. For example, sacrificing some accuracy for speed – in our ‘Conference Valet’ application, the attendees utterances would need to be analyzed in short quick bursts requiring a moderate level of accuracy in order to extract the intent. Let’s now flip it – sacrificing speed for accuracy with a ‘Doctor/Patient video consultation’ application where the transcripts are needed for compliance and accuracy is critical. 

Beyond speed and accuracy, there are value add-on features - take for example, Mod9’s - cloud-based service called ‘ReMeeting.’ They specialize in not only high levels of accuracy but also speaker separation and searchability - powerful features that can help innovate specific applications. Last but not least, the ability to train or tune the speech analytics engine 'out of the box' to better serve the specific application. For instance, a voicemail application with email transcriptions almost always contains a call back telephone number which should be interpreted as an integer rather than words ('my number is 7169.....' vs 'my number is seven one six nine....')

In the end, the best speech analytic engine will *depend* on the *use case* so be sure to compare the strengths (and weaknesses) against your *application requirements* before making a decision. 

//Vince - @vfpuglia

Liked this post? Get more content like this right to your inbox. Subscribe to  the blog.

Topics: Artificial Intelligence