This merged direction connects speech intelligence with LLM efficiency. Instead of crowding the visual with equations, the hero now works like a motion scene: waveform, feature activity, memory bank, context graph and compressed output all move in one continuous flow.
VowelVC emphasizes vowel-centric acoustic cues and uses incremental memory learning for few-shot speaker adaptation, aiming to retain identity with limited target data.
Graph-aware prompt compression represents reusable reasoning units as a structured graph and applies reinforcement learning to keep compact but informative contexts for black-box LLM inference.
Across speech and language, the shared idea is selective efficiency: represent the right structure, store it intelligently, and use only the parts that change the outcome.
The VowelVC paper and the LLM prompt-compression paper now live together here as one coherent direction inside the portfolio.
How little target data can a model see and still adapt effectively without forgetting useful prior knowledge?
How much context can be removed before reasoning quality actually degrades in practice?
Can efficient AI remain interpretable when it works through memory selection, graph structure and learned policies?
Research, engineering, or a stubborn technical problem — send it over.