EVENT DOSSIER
On September 7, 2026, arXiv released GEPARD, a streaming text-to-speech model based on large language model (LLM) cores, designed for real-time conversations. The model uses a single-decoder architecture, jointly training text and audio embeddings, and employs a FSQ neural encoder to generate waveforms in blocks. Its core design principle is to not modify the computational core of standard LLM engines (such as vLLM). In single-stream inference, GEPARD’s real-time factor is approximately 0.067, achieving 15 times the speed of real-time processing; when supporting 256 concurrent streams on a single server-level GPU, the overall acceleration ratio can reach approximately 204 times. The paper also details methods for diagnosing and mitigating “short registration” failure modes in autoregressive speech decoders, as well as the technical details of freely guiding two classifiers into a single weight using direct preference optimization (DPO).
GEPARD GEPARD 生成式韵律感知自回归 TTS 模型发布
GEPARD 是一款基于 LLM 骨干的流式文本转语音模型，旨在实现实时对话。该模型采用单一解码器架构，将文本与音频嵌入联合训练，并通过 FSQ 神经编码器逐块生成波形，核心设计原则是不修改标准 LLM 引擎（如 vLLM）的计算内核。在单流推理中，其实时因子约为 0.067，速度约为实时的 15 倍；在单个服务器级 GPU 上支持 256 个并发流时，整体加速比可达约 204 倍。论文详细阐述了 vLLM 原生服务的系统级解决方案、针对自回归语音解码器“短注册”故障模式的诊断与缓解方法，以及通过直接偏好优化（DPO）将两路分类器自由引导蒸馏为单路权重的技术细节。