Roadmap for Evaluating a Voice Biometric System

Regardless of your industry or your company’s specific objectives, testing and deployment of a voice biometric solution can be easy and fast… as long as you rely on proven methodology. In this emerging market full of newcomers, PerSay stands out with its long track record and expertise in global deployment of voice biometrics. While many speech security solutions are still in the pilot stage, or are highly customized to fit the needs of a particular niche, PerSay has deployed over 60 different systems worldwide, with customers acknowledging seamless integration with existing systems and ease of deployment. 

In addition to this unmatched expertise and documented customer satisfaction, PerSay’s set of Performance Evaluation Tools makes it easy to quickly and effectively evaluate a voice biometric system prior to implementation, following these simple steps:

• Evaluation Planning
  Define goals; coordinate resources, processes and responsibilities; develop schedule and select pass phrase—either common to all speakers or unique to each

• System Installation and Setup
  Integrate voice biometric system with your IVR, a hosted service, or use one of PerSay's enabled voice platforms to verify functionality of all components. For a VoiceXML platform, use PerSay's VoiceXML sample scripts.  

• IVR Call Flow
  Design two simple IVR call flows for acquiring a single utterance of audio from each speaker, and for creating a voiceprint or performing verification. The verification call flow should contain an authenticity feedback—notifying the system when an impostor session takes place. This capability enables the system to calculate the evaluation error rates. Note: If the evaluation will be performed with pre-recorded files, you can use PerSay’s Performance Evaluation Tool (PET) to send files or directories to the system… or you can use PerSay’s scripting tool.

CLICK TO ENLARGE

• Calibration
  Calibration is the process of optimizing system performance by building a Background Model that represents the audio characteristics of your specific operating environment. This is done by collecting at least 100 utterances from different speakers, creating the Background Model, and uploading it to the system.
  

  Note: calibration should precede system evaluation by end-users.

• Caller Enrollment
  At least 50 new participants call the IVR enrollment line, identify themselves, and create their voiceprints. If you’re using unique pass phrases, produce a written document containing speaker IDs and pass phrases for each caller.

• Accuracy Testing
  Authentication: Callers dial up system, identify themselves, and say their pass phrases. Each caller should perform at least five verifications, on different days, from different channels. Verification scores are produced and stored in the system.

  Imposters: When using a common pass phrase for all speakers, imposter testing can be done offline. In the case of unique pass phrases, give callers the IDs and pass phrases of other speakers (same gender) enrolled in the system, and request that they call in as “imposters.” It is recommended that this be done by allocating a different line for imposters (the voice application running on this line will then feed this information back to the system). At least five imposter verifications should be attempted.

• Analysis and Summary
  When the evaluation process includes tuning, decision thresholds are set for acceptance and rejection of claimed identity at an optimum working point that’s based on an estimated Decision Error Tradeoff curve.

In-depth analysis of evaluation results leads to a summary report that includes: error analysis (false accept and false reject performance curves), specific values based on selected upper and lower decision thresholds, and recommendations for deployment. Note: Equal Error Rate (EER) is when the false accept rate equals the false reject rate.

Success Stories:

Here’s a snapshot of a large-scale deployment that reinforces the ease and speed of the process outlined above:

Bell, the largest communications provider in Canada, has made a spoken pass phrase the universal identifier for customer support across five service segments (wireline, wireless, high-speed Internet, digital TV and VoIP). Voluntary enrollment is offered to customers as a more convenient alternative to an array of PIN-based services and manual authentication methods, with voice biometrics supporting the company’s ongoing commitment to protect privacy as it offers increasingly complex and varied services. PerSay VocalPassword was selected for the Bell project in September 2006. Installation, integration by IBM Global Business Services, and testing followed quickly, and beginning in March, Bell’s Voice Identification Service was rolled out, one area code at a time, to all wireline and wireless customers in Canada. As of August 20, 2007, Bell reported more than 275,000 voluntary enrollments.