LLaMA-Omni2: LLM-based Real-time Spoken Chatbot with Autoregressive Streaming Speech Synthesis

TL;DR

Real-time spoken interaction with large language models remains challenging due to cascading errors and latency. LLaMA-Omni 2 presents a modular SpeechLM that combines Whisper-based speech understanding with an autoregressive streaming TTS pipeline and a flow-based vocoder to achieve real-time speech generation with 0.5B–14B parameter LLMs. The approach uses a two-stage training regime on 200K multi-turn dialogues and demonstrates strong performance on spoken question answering and speech instruction following, outperforming prior SpeechLMs such as GLM-4-Voice and LLaMA-Omni, while maintaining latency below ~600 ms. These results indicate a data-efficient path to robust, real-time speech dialogue with large-scale models, with potential extensions to emotion and dialect-aware synthesis.

Abstract

Real-time, intelligent, and natural speech interaction is an essential part of the next-generation human-computer interaction. Recent advancements have showcased the potential of building intelligent spoken chatbots based on large language models (LLMs). In this paper, we introduce LLaMA-Omni 2, a series of speech language models (SpeechLMs) ranging from 0.5B to 14B parameters, capable of achieving high-quality real-time speech interaction. LLaMA-Omni 2 is built upon the Qwen2.5 series models, integrating a speech encoder and an autoregressive streaming speech decoder. Despite being trained on only 200K multi-turn speech dialogue samples, LLaMA-Omni 2 demonstrates strong performance on several spoken question answering and speech instruction following benchmarks, surpassing previous state-of-the-art SpeechLMs like GLM-4-Voice, which was trained on millions of hours of speech data.

Figures (1)

• Figure 1: Left: Model architecture of LLaMA-Omni 2. Right: Illustration of the two-stage training strategy.