← Back to Paper Radar

Open Full-duplex Voice Agent with Speech-to-Speech Language Model

Edresson Casanova, Chen Chen, Ke Hu, Ankita Pasad, Elena Rastorgueva, Seelan Lakshmi Narasimhan, Slyne Deng, Ehsan Hosseini-Asl, Piotr Żelasko, Valentin Mendelev, Subhankar Ghosh, Yifan Peng, Zhehuai Chen, Jason Li, J. Balam, Vitaly Lavrukhin, Boris Ginsburg
Automatic Speech Recognition & Understanding | 2025
The paper introduces a data-efficient method for converting any pretrained text-only LLM into a full-duplex, end-to-end speech-to-speech (S2S) model capable of simultaneously listening and speaking, paired with a fully open-source pipeline for low-latency voice agent deployment.

Problem Statement

Human-like conversational voice agents require full-duplex interaction (simultaneous listening and speaking), which is difficult to achieve with traditional cascaded ASR→LLM→TTS pipelines due to latency and rigid turn-taking. Existing end-to-end S2S models that support duplex behavior typically demand extensive, costly speech-text pretraining, restricting access to well-resourced labs. This work targets that accessibility gap by offering a generalizable, low-data, fully open-source alternative.

Key Novelty

  • A data-efficient adaptation method that converts any standard text LLM into a full-duplex S2S model without extensive speech-text pretraining
  • A modeling approach enabling simultaneous listen-and-speak behavior directly within the LLM, rather than through separate turn-based modules
  • An integrated open-source stack combining modeling, inference optimization, and serving for real-time, low-latency voice interaction
  • A generalizable framework designed to work across different base LLMs rather than being tied to one specific architecture

Evaluation Highlights

  • System demonstration of a working low-latency, full-duplex voice agent built end-to-end from open-source components
  • Qualitative comparison highlighting reduced speech-text pretraining data requirements relative to conventional S2S approaches

Signal Assessment

5/10 The work delivers a practical, valuable engineering and open-source contribution—efficient LLM-to-S2S conversion and a full deployable stack—but builds on existing full-duplex S2S concepts (e.g., Moshi-style models) rather than introducing a fundamentally new capability.

Methodology

  1. Start from a pretrained, standard text-only LLM as the reasoning/generation backbone
  2. Apply a lightweight, data-efficient adaptation procedure to add speech input/output modeling without full-scale speech-text pretraining
  3. Introduce a full-duplex architecture allowing concurrent processing of incoming audio and generation of outgoing speech tokens
  4. Optimize inference (e.g., streaming decoding, caching) to minimize end-to-end response latency
  5. Package the model with open-source serving infrastructure to deploy a real-time conversational voice agent

System Components

Text LLM backbone

Any standard pretrained language model that provides the core reasoning and generation capability

Speech tokenizer/codec

Converts raw audio into discrete tokens and back, bridging speech and the LLM's token-based interface

Full-duplex modeling mechanism

Enables the model to process incoming speech and generate outgoing speech simultaneously, supporting natural, interruption-tolerant dialogue

Data-efficient adaptation method

Fine-tuning/adaptation technique that imparts speech capabilities to the LLM without requiring extensive speech-text pretraining corpora

Inference optimization layer

Techniques (e.g., streaming, efficient decoding) that reduce latency to enable responsive, real-time conversation

Open-source serving pipeline

Deployment infrastructure that assembles the model and optimizations into a usable, low-latency voice agent

Results

Aspect Conventional S2S Approach This Paper Improvement
Speech-text pretraining data needed Extensive/large-scale Minimal (data-efficient) Substantially reduced data requirement
Duplex capability Often half-duplex/turn-based Full-duplex (simultaneous listen & speak) More natural, human-like conversation
Latency High (cascaded pipeline overhead) Low-latency, real-time Improved responsiveness
Accessibility/openness Often closed or resource-intensive Fully open-source stack (model, inference, serving) Lower barrier to entry for practitioners

Key Takeaways

  • Practitioners can turn existing open-source text LLMs into speech-capable conversational agents without collecting or training on massive speech-text corpora
  • Full-duplex (simultaneous listen/speak) behavior can be engineered into standard LLM architectures rather than requiring bespoke speech-native models built from scratch
  • The full open-source release (modeling + inference optimization + serving) offers a ready blueprint for deploying production-grade, low-latency voice agents
  • This approach could accelerate adoption of natural, human-like voice AI by smaller teams lacking access to large proprietary speech datasets or infrastructure

Abstract

We present the system demonstration and opensource code release of a novel, data-efficient framework that converts any standard text Large Language Model (LLM) into a full-duplex end-to-end (E2E) speech-to-speech (S2S) model, for building conversational voice agents. Our new modeling method enables any LLMs to simultaneously listen and speak without requiring extensive speech-text pretraining. Moreover, we demonstrate how to put together a low-latency and full-duplex voice agent with open-source modeling, inference optimization, and serving solutions. This work significantly lowers the barrier to entry for developing low-latency, human-like voice agents by providing a generalizable, end-to-end solution built on open-source technologies.

Generated from available metadata and abstract on 2026-07-14 using Claude.