Paper · arXiv 2503.04724
LLMVoX: Autoregressive Streaming Text-to-Speech Model for Any LLM
Sambal Shikhar, Mohammed Irfan Kurpath, Sahal Shaji Mullappilly, Jean Lahoud, Fahad Khan, Rao Muhammad Anwer, Salman Khan, Hisham Cholakkal
LLMVoX is a lightweight, LLM-agnostic TTS system that decouples speech synthesis from model processing, ensuring high-quality speech with low latency and supporting seamless multimodal dialogue.
Abstract
Recent advancements in speech-to-speech dialogue systems leverage LLMs for multimodal interactions, yet they remain hindered by fine-tuning requirements, high computational overhead, and text-speech misalignment. Existing speech-enabled LLMs often degrade conversational quality by modifying the LLM, thereby compromising its linguistic capabilities. In contrast, we propose LLMVoX, a lightweight 30M-parameter, LLM-agnostic, autoregressive streaming TTS system that generates high-quality speech with low latency, while fully preserving the capabilities of the base LLM. Our approach achieves a significantly lower Word Error Rate compared to speech-enabled LLMs, while operating at comparable latency and UTMOS score. By decoupling speech synthesis from LLM processing via a multi-queue token streaming system, LLMVoX supports seamless, infinite-length dialogues. Its plug-and-play design also facilitates extension to various tasks with different backbones. Furthermore, LLMVoX generalizes to new languages with only dataset adaptation, attaining a low Character Error Rate on an Arabic speech task. Additionally, we have integrated LLMVoX with a Vision-Language Model to create an omni-model with speech, text, and vision capabilities, without requiring additional multimodal training. Our code base and project page is available at https://mbzuai-oryx.github.io/LLMVoX .