Score: 0

LLMVoX: Autoregressive Streaming Text-to-Speech Model for Any LLM

Published: March 6, 2025 | arXiv ID: 2503.04724v1

By: Sambal Shikhar , Mohammed Irfan Kurpath , Sahal Shaji Mullappilly and more

Potential Business Impact:

Lets computers talk and understand like humans.

Business Areas:
Speech Recognition Data and Analytics, Software

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 .

Country of Origin
🇦🇪 United Arab Emirates

Page Count
12 pages

Category
Computer Science:
Computation and Language