Abstract
Current voice-activated conversational agents rely on sequential, text-based processing pipelines that introduce latency and fail during mid-utterance interruptions. A computational architecture is disclosed that replaces traditional speech recognition pipelines with a streaming native-audio large language model. Acoustic tokens are processed directly to maintain a real-time computational graph of an acoustic trajectory. Streaming latent diarization is employed natively within the model to mathematically segregate intersecting vocal trajectories. Parallel isolated execution branches are dynamically created when biometric shifts are detected, allowing background noise to be masked. Structured tool-calls are generated speculatively by an execution head before end-of-speech markers are reached. Conversational intent execution is thereby decoupled from strict acoustic endpoints, resolving mid-utterance multi-speaker cross-talk and reducing conversational latency.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yakar, Tamar and Labzovsky, Ilia, "Anticipatory Action Fulfillment in Conversational Agents via Speculative Native-Audio LLMs and Streaming Latent Diarization", Technical Disclosure Commons, (July 16, 2026)
https://www.tdcommons.org/dpubs_series/10990