New Data-Efficient Technique Enables Full-Duplex Voice Agents Using Frozen LLMs
In short: Researchers have developed a method to adapt general-purpose, frozen LLMs into full-duplex voice systems using only 97 million trainable parameters and 30 GPU-hours of training. The system functions as a thinker-talker cascade with an 80 ms frame clock, allowing for realistic turn-taking, interruptions, and backchanneling without massive speech-data pretraining.
This summary was generated automatically by AI from arXiv cs.CL's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Introduced a thinker-talker cascade architecture
- 2Uses a 80 ms frame clock for token emission
- 3Frozen base LLM with 97M trainable parameters for adapters
- 4Four-stage curriculum training process
Why it matters
This advancement drastically reduces the compute requirement for building state-of-the-art full-duplex voice agents. By avoiding the need for millions of hours of speech training, developers can now achieve competitive performance in turn-taking and spoken-task accuracy much faster.
What it means for AI agents and contact centers
For your team building voice agents, this architecture offers a template to achieve low-latency, human-like turn-taking (interrupts and backchanneling) by augmenting existing LLM APIs rather than training custom speech-native models. Compare this approach’s latency and accuracy in turn-taking against your current pipelines.
🧪 Worth Testing
The methodology is highly relevant for optimizing agent responsiveness and natural conversational flow without the resource-heavy overhead of traditional voice-model training.
Sources
- arXiv cs.CLOfficialPrimary sourceOriginal article →„An LLM Is All You Need: Data-Efficient Adaptation to Full-Duplex, Word-Level Spoken Dialogue“13 Nov 2026, 07:00Licence: arXiv metadata CC0; papers under their own licences · our summary (content changed)
- Published by source
- 13 Nov 2026, 07:00
- Found by our system
- 13 Nov 2026, 09:45
- Summary generated
- 13 Nov 2026, 09:45
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy