LG Electronics’ flagship laptop brand ‘Gram’ has integrated voice recognition technology from South Korean AI startup ActionPower. This enables Gram users to convert recorded audio files and real-time speech into text and receive automatic summaries without an internet connection.
ActionPower announced on the 29th that it has supplied its on-device speech-to-text (STT) model for LG Gram’s AI note application, ‘LG Gram Note.’ According to industry sources on the 28th, the technology will be rolled out sequentially through updates starting with 2026 Gram models.
ActionPower, which operates the AI-based dictation service ‘Daglo,’ successfully compressed its server-based speech recognition engine to a level capable of running in a laptop environment through this collaboration. The key challenge was delivering stable performance with limited computing resources, and the company emphasized that it achieved recognition rates equivalent to its server-based model.
The AI package installed in LG Gram Note includes: an on-device STT model that operates without internet connectivity; an automatic speaker separation function; and a summarization-specialized on-device small language model (SLM). This allows users to process features such as script highlighting by playback segment, script editing, bookmarks, and automatic summarization entirely on the device.
Notably, speaker separation accuracy exceeded 90% based on publicly available application programming interface (API) model benchmarks. ActionPower also explained that its newly developed summarization-specialized SLM boosted summarization performance by up to 36% while maintaining a parameter count of under 2 billion.
The greatest strength of on-device AI is its ability to reduce the potential for external leakage of personal information or corporate secrets, as all data is processed directly on the terminal rather than on servers. Additional advantages include unrestricted use in offline environments and reduced network traffic and cloud inference costs.
Daglo is a service offering work-specialized AI features including dictation, automatic summarization, meeting minutes creation, AI search, and presentation generation. It has recorded over 2.2 million cumulative downloads, primarily among university students and office workers. The service is recognized for applying large language models (LLMs) optimized for specific tasks such as meeting summary and information extraction, achieving equal or superior performance at lower inference costs compared to general-purpose models, while also resolving hallucination issues through fine-tuning.
An LG Electronics representative stated, “We are collaborating with ActionPower to provide on-device AI,” adding, “Updates are being provided to select products and customers.” LG Electronics reportedly conducted thorough proof-of-concept (PoC) testing for this technology integration.
With this deployment, ActionPower has expanded the scope of its AI model applications from existing cloud and on-premise environments to personal device territory. ActionPower co-CEOs Cho Hong-sik and Lee Ji-hwa said, “It is significant that we have implemented speech recognition and summarization AI — which previously remained in server environments — on personal devices without internet connectivity.” They added, “Based on our ultra-lightweight AI technology, we plan to expand our on-device business with the goal of realizing ‘AI for each individual,’ where dedicated AI trained on users’ data and usage habits runs directly on various devices including PCs and smartphones.”