“We’re developing our own STT (Speech-To-Text) model that converts voice into text and will likely be available throughout the yr.
“We’re upgrading ‘Collabo’ beyond the STT (speech-to-text) service for organizing meeting minutes to a company productivity tool. To this end, we’ve developed our own small language model (sLM). The model performance has also been top-notch.”
ReturnZero (CEO Lee Cham-sol) is a voice recognition AI company well-known for its meeting minutes writing AI service ‘Collabo’ and app ‘Bito’ that converts recorded phone calls into text.
Based on the Korean language data gathered thus far, it has presented accurate voice recognition technology through an engine that analyzes voices and separates speakers, and an engine that converts phone call voices into text in real time. In January, it was also announced that the cumulative voice processing time exceeded 15 million hours in 3 years and 10 months. Which means that it’s the most effective within the country on this field.
Here, we’ve developed a small language model (sLM) to further enhance collaboration technology this yr.
Kim Hyo-shin, COO of ReturnZero, explained the event background, saying, “There was a requirement from corporations to deliver executive meetings with appropriate motion items to every department, and the summary function that matches security issues and context was vital.”
He continued, “Corporations which have introduced the actual model are showing favorable reviews of the motion item function,” and emphasized, “Collab can contribute to increasing corporate work productivity by summarizing meeting content and sharing motion items through continuous relearning and prompting that suits various industries and meeting contexts.”
It was also revealed that the extent of completion for its own model was high.
CTO Jung Joo-young said that sLM ranked 2nd within the W&B Tiger Leaderboard for parameters under 10B. As well as, sLM achieved 1st place within the ‘LogicKor’ Leaderboard, which measures the multi-disciplinary considering ability of Korean language models. This benchmark is a leaderboard that has recently turn into popular amongst developers.
CTO Jeong said, “SLM is a gateway to developing a big language model (LLM).” He added, “The utility felt by actual consumers is more vital than simply comparing model performance.”
Noh Jeong-gyu, head of research, said, “When using existing general-purpose models, meeting summaries were simply provided through prompting, but with SLM, fine-tuning and repair could be done in keeping with a transparent purpose.” He also said, “The intention is to cut back response delay and lower costs by miniaturizing the model.”
As well as, it was revealed that “currently, major financial institutions are positively reviewing the introduction of collaboration.” Return Zero has been working with the financial sector, comparable to providing voice recognition solutions for the establishment of Shinhan Financial Group’s AI Contact Center (AICC) earlier this yr.
“High-precision voice recognition, re-learning, and keyword boosting enable understanding of complex product names of monetary institutions,” he said. “Continuous re-learning improves the performance of the STT (speech-to-text) engine, allowing convenient use without having to construct multiple solutions.”
In addition they revealed plans so as to add multilingual features.
CTO Jeong said, “We’re currently developing real-time interpretation, following the function that enables attendees to translate into their preferred language after a gathering,” and “We are going to repeatedly add functions essential to enhance corporate productivity based on customer requests.”
Specifically, it was reported that when meeting minutes are written and shared through collaboration, accessibility to details about other teams or projects increases. It was explained that the main focus was on improving user convenience by supporting integration with collaboration tools comparable to Slack, Salesforce, and Zapier.
CTO Jeong said, “This was possible since the essential idea when ReturnZero was founded was business coaching and sales coaching.”
“It was planned as a tool to supply feedback during sales coaching within the sales department,” he said, emphasizing that “Collabo is scaling up from a gathering minutes function to a company productivity tool.”
Reporter Park Soo-bin sbin08@aitimes.com
