EEG Video Integrates Core Models Feature to Lexi Automatic Captioning Service

EEG Video has announced availability of Core Models for its popular Lexi Automatic Captioning solution. The new offering represents a powerful expansion of EEG’s robust Topic Models system for accurately displaying custom vocabulary and phrases.

With the addition of Core Models, Lexi users and their audiences will experience even higher captioning accuracy for a wider variety of content. Core Models further expand the usability and quality of affordable AI captioning for live broadcasting, live streaming and events, videoconferencing, and even VOD programming.

The Core Models system builds on the usefulness of EEG’s Topic Models feature. This breakthrough technology enables Lexi to recognize topics, immerse itself in distinctive vocabulary, and observe context through the absorption of relevant web data unique to each implementation. Topic Models enable Lexi to perform in real-time with a high degree of accuracy by better addressing the poor recognition of topic-specific vocabulary often displayed by previous speech-to-text captioning systems. This can include less common proper nouns, such as names of people, places, or products. Topic Models also address entire phrases, jargon, vocabularies or speaking styles that would be typical in one context, such as a baseball telecast, but unusual in another context, such as a medical imaging conference.

Additionally, people and phrases in the news can change rapidly. However, many systems update basic vocabulary infrequently–on a quarterly basis or less–and still do not necessarily weigh recent new developments highly enough compared to older sources of training data.

Phrases like “coronavirus” and “COVID-19,” for example, have been used continuously in TV news coverage for more than six months, yet several off-the-shelf commercial speech-to-text engines still do not recognize these phrases, providing poor phonetic substitutions, like “culvert” for the previously unknown word “COVID.” EEG recognizes that eliminating such discrepancies are critical for breaking news broadcasters to assess the viability and credibility of AI captioning, as well as achieving improved engagement and satisfaction for audiences.

With more than three years of experience in providing AI captioning for a myriad of events and media, EEG’s AI Team has distilled some of the most common vocabulary training cases into a set of Core Models, a subset of Topic Models maintained by EEG experts and available to all Lexi customers. Customers can also build their own individualized Topic Models on top of an EEG Core Model, creating multiple layers of accuracy-enhancing customization. As EEG continues to evolve and add to the Core Model, the derived individual customer models are also automatically updated to merge individual and Core changes.

Current EEG Core Models exist in English for:

• Headline News (United States-focused): more than 15,000 entities and phrases
• Sports: Baseball (MLB-focused): more than 10,000 entities and phrases
• Christian Broadcasting: more than 15,000 entities and phrases
• Legislative and Municipal Sessions: more than 1,000 entities and phrases
• Weather (United States focused): more than 1,000 entities and phrases

The list of Core Models is expected to grow in the coming months, which will include entries in additional languages.

“It only takes one look at off-the-shelf speech-to-text transcripts to realize that customization and updates are a critical task, but one that many end users are limited in time and experience to effectively perform on their own,” says Bill McLaughlin, VP of Product Development for EEG Video. “EEG’s experts have the deep data curation skills needed to update AI models effectively and efficiently. This leverages the shared needs of our many AI captioning customers to produce Core Models, dramatically reducing the customization requirements of individual customers in common applications.”