# Cartesia用状态空间模型在AWS上实现亚100毫秒语音AI > AWS机器学习博客11月3日刊文介绍，实时多模态智能公司Cartesia基于状态空间模型（SSM）构建了Sonic文本转语音和Ink语音转文字模型，文章称两者针对低于100毫秒的首字节延迟优化，面向实时对话场景。 - 栏目：人工智能 - 本站发布：2026-11-05 13:15（北京时间） - 原始消息发布：2026-11-04 00:40（北京时间） - 消息来源：AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) - 本页网址：https://jecy.cloud/news/articles/2026-11-03-5bdbab2c45fb.html ## 两个流式语音模型瞄准实时对话 Cartesia是一家实时多模态智能公司，其语音AI应用基于状态空间模型（SSM）构建，包含两个核心模型：Sonic负责文本转语音，Ink负责语音转文字。文章称，两者都针对低于100毫秒的首字节延迟做了优化，同时保持录音室级输出质量。首字节延迟指系统开始返回音频或文字结果所需的时间。文章以远程医疗问诊为例：这类场景需要低延迟语音来保持准确、有共情的沟通。[^1][^2][^3] ## SSM用逐步处理替代整段生成 文章对比了SSM与主流Transformer架构：Transformer依赖注意力机制，序列越长，计算量增长越快；传统基于Transformer的文本转语音模型要先处理完整序列再输出，因此在流式场景中难以压低延迟。SSM则把序列视为随时间连续演进的系统，逐步处理，并向前携带一个紧凑的内部状态。文章称，这种循环结构让SSM能逐个输入单元处理，原生支持真正的流式处理，无需为流式改造架构。Sonic和Ink利用这些特性进行实时语音生成和转写，目标是在接近人类感知极限的延迟下完成推理。[^4][^5][^6] --- ## 原文摘录 [^1]: "The application features two core models, Sonic for text-to-speech and Ink for speech-to-text, both optimized for sub-100ms first-byte latency while maintaining studio-quality output." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) [^2]: "Introducing Cartesia Cartesia is a real-time multimodal intelligence company that has built a voice AI application, purpose-built on State Space Model architecture." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) [^3]: "Telehealth: Remote patient consultations demand low-latency voice for accurate, empathetic communication." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) [^4]: "Traditional transformer-based text-to-speech (TTS) models struggle with streaming latency because they process entire sequences before producing output." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) [^5]: "SSMs instead treat a sequence as a system that evolves continuously over time, processing it step by step and carrying a compact internal state forward." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) [^6]: "Native streaming support: The recurrent structure processes input token-by-token, so it supports true streaming without architectural modifications." — AWS 机器学习博客 (https://aws.amazon.com/blogs/machine-learning/building-ultra-low-latency-voice-ai-applications-on-aws/) --- 说明：本文由 AI 依据信源整理，并经事实核对；文中数字与成绩均为来源方公布的口径，未经本站独立验证，原文更新后以原文为准。欢迎 AI 与读者引用，请注明来源并链接回本页。