Ground Truth Data Collection Services for AI – National Speech Data Collection

BACKGROUND

A popular social media company was preparing to enter the in-home smart-hub market in the fourth quarter of 2018. Already lagging behind the introductions of other smart-hub devices, the company understood that its success relied upon targeting a narrow set of features for the device (specifically, calling) and gathering a large and diverse set of speech utterance data to ensure that the device would recognize and respond accurately when prompted.

CHALLENGE

Early testing and data confirmed that the device was struggling with strong accents and varied tones or pitches. The company needed more data to build the patches necessary to improve its voice-recognition software. To capture these speech tones and patterns management decided to leverage its diverse employee population to “dogfood” the technology for capturing speech tones and patterns. With the product launch just months out from the start of data collection, the company was compelled to short-list vendors who could mobilize teams across the country to deliver a “first-in-kind” data collection program within a very short timeframe. Q Analysts quickly rose to the top of the list.

SOLUTION

Q Analysts provides a full suite of Ground Truth Data Services, including data collection, ingestion, and automated and manual tagging to create algorithm-ready quality data optimized for artificial intelligence (AI) and machine learning (ML). The company’s complete methodology initiates, creates, and delivers on a variety of data collection initiatives for areas such as speech, humans, and spaces.

With its expertise in speech data capture, such as accents, dialects, speech patterns, cadence, and other speech nuances and behaviors, Q Analysts was able to quickly scope out and deliver the teams necessary to target the company’s preferred employee locations across the country. Q Analysts partnered with the client to develop a strategy and deploy the mobile teams, staffed with leads and moderators, to conduct test sessions with the company’s full-time employees.

The biggest challenge of this project was enrolling participants. To create interest, the project team developed a multi-faceted marketing program that included internal social media promotion, free swag, targeted flyer placement, and live interaction with employees during high traffic mealtimes. Once employees had signed-up to “dogfood,” Q Analysts fine-tuned its communication strategy to ensure leads would honor their commitment and show up for their scheduled sessions. The team used a “help-chat” model to engage and remind employees of available time slots. Proactively working with internal stakeholders from the engineering team, Q Analysts was able to implement several innovative program efficiencies. Cross-training test session teams on best-practices and establishing a steady cadence of reporting allowed for optimal knowledge collection and sharing between offices, ensuring client visibility and engagement throughout the project.

RESULTS

Setting its own benchmarks, the Q Analysts team delivered statistically significant results in all phases of the program: program promotion, sign-ups, conversion to sessions, and utterances captured. Based on the initial success of the program, the social media company is considering the possibility of replicating the dogfooding program internationally and for other client teams with similar data-collection needs.

Program Analytics At-A-Glance

  • -Successfully deployed mobile dogfooding teams across select company’s campuses nationwide
  • -Leveraged in-house and ad-hoc promotional materials to conduct 90+ live marketing events
  • -Utilized multi-channel strategy to secure 4,000+ sign-ups in two months
  • -Optimized post-launch communication process, driving leads from “sign-up” to “session,” resulting in 65% programmatic conversion rate
  • -Operationalized system to efficiently conduct over 2,500 successful dogfooding sessions resulting in 115,000 utterances over a two-month period.

Comments are closed.