SpeechCycle Introduces Vendor-Independent Grammar Factory

SpeechCycle today launched nRich Grammar Factory—a new subscription service aimed at improving, enriching, and optimizing the speech application and interactive voice response (IVR) system caller experience.

According to Roberto Pieraccini, chief technology officer at SpeechCycle, the company’s message with this release is a simple one:

“Don’t worry about your grammars anymore.  Just give us your logs and we will generate grammars for you and we do that so efficiently and such an automated way that it is going to continuously feed your grammars.  The more data we have the more grammars we generate.”

To generate grammars, Pieraccini says, nRich Grammar Factory utilizes three primary building blocks:

• Secure IVR and Application Data Extraction, which captures caller interaction data and securely transmits it to the SpeechCycle Ops Center;
• Caller Data Tagging Operations: which processes caller interaction data and tags that data to IVR and application business logic, as a preparatory step to the grammar creation process; and
• Data Conditioning and Grammar Creation, which produces new grammars for the targeted IVR and/or applications, which are then made available to customers and developers through a secure Grammar Factory Portal.

“In order to be able to build grammars automatically, we need to tag all the utterances of the recordings that we get with a correct word,” Pieraccini says. “Once we have this tagged data there automated processes that first of all condition the data—clean up the data, make it consistent, remove the inconsistencies, look at the issues with the data—and then create grammars and do all the testing for the grammars, all in an automated fashion.”

According to Pieraccini a major impediment to wider adoption of speech-enabled IVRs is what he calls the “mystery of the science”—a mystery that nRich Grammar Factory helps to demystify.

“Every IT developer can build a speech IVR,” he says. “But when they have to go build grammars, tune the grammars, they typically have to go to a professional service company…and have them tune the grammars.”

Pieraccini says this tuning process often lacks visibility, ignores continuous testing, and results in companies paying for grammars that “someone claims work better.”
“This is a revolutionary step toward making speech applications as easy to build as websites,” he says.  “This is the first step.” 

According to Pieraccini, the offering reduces dependence on proprietary vendor upgrades and professional services—something that will save companies money.

“The system fixes itself.  It’s moving in the direction of what people call autonomous systems.  They learn—it keeps learning.  But it saves money also because it gets better caller experience and presumably higher automation.”    

And an automated process, Pieraccini says, can look at significantly more data than a manual process.

“What we use, rather than rule-based grammars, we use call statistical grammars,” he adds.  “We can feed grammars continuously. What limits us is the release process of the customer…We could continuously release grammars.  So we decide with the customer the frequency of the [release] date.”