[SecureMemo] Release of next-gen speech recognition model 'shirushi-2.0'. Combining with the existing 'Personalized Correction Function' to create a faster, smarter, and organization-adaptive AI meeting minute tool.

Nishika Inc. (Headquarters: Shinagawa-ku, Tokyo; CEOs: Tatsuro Yamashita, Hiroyuki Matsuda; hereinafter 'Nishika'), a company engaged in AI product and consulting businesses, has released 'shirushi-2.0', the next-generation model for its 'shirushi' series of speech recognition AI installed in the AI meeting minute tools 'SecureMemo' and 'SecureMemoCloud'.

With this release, SecureMemo's AI meeting minutes now feature a two-layer structure: a layer for 'fast, smart, and accurate transcription' (shirushi-2.0) and a layer for 'personalization that learns organization-specific vocabulary' (Personalized Correction Function). The goal is to achieve the level of detail where 'meeting content is turned directly into meeting minutes,' which general-purpose speech recognition AI could not previously provide.

Background: The demands for AI in meetings: 'Speed,' 'Accuracy,' and 'Organizational Adaptation' The practical value of speech recognition AI cannot be measured solely by general benchmark figures. From our work with many client companies, we repeatedly heard three voices: 'I want to check minutes immediately after a meeting' (Speed), 'Correcting misconceptions takes too much time' (Accuracy), and 'It doesn't learn our proprietary terms' (Organizational Adaptation). Based on the hypothesis that metrics like CER and WER do not align with field experience, Nishika has proceeded with development in two directions.

Features of 'shirushi-2.0': Fast and Intelligent Transcription 1. Over four times faster processing speed: In business meeting audio of about one hour, it achieves an average speed increase of more than four times compared to the previous 'shirushi-1.5' model. 2. Improved Noun Recognition Accuracy (noun-WER): By excluding punctuation and transcription variations, we achieved an improvement of several points in noun-WER, which evaluates only the 'nouns' that affect the semantic value of the minutes.

Examples of Speech Recognition Improvements (Excerpt) - Qualification names and Fiscal years: Construction Management Engineer (First Class), 2024 fiscal year (Corrected). - Accounting/Personnel: Closing settlement, strict observance (Corrected). - Project management: Process chart, rough estimate, oral (Corrected). - IT/Katakana terms: Drag-and-drop, explanation video, field experiment (Corrected).