Acoustic-to-Text KV Compression for Full-Duplex Speech Models
An acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval, which reduces peak streaming KV-cache size and improves transcription, temporal question answering, and summarization over the baseline.
Ye-Jin Lee, Seungbeom Kim, Yong-Ha Lee et al.
· 0 citations