LLMVoX: Autoregressive Streaming Text-to-Speech Model for Any LLM
This addresses the need for efficient, high-quality speech synthesis in conversational AI without compromising LLM performance, though it appears incremental as it builds on existing TTS and LLM integration methods.
The paper tackles the problem of speech-to-speech dialogue systems being hindered by fine-tuning, computational overhead, and text-speech misalignment, proposing LLMVoX, a lightweight autoregressive streaming TTS model that achieves a significantly lower Word Error Rate compared to existing speech-enabled LLMs while preserving LLM capabilities.
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 .