[ ● ] Sonari Labs · Established April 2026
[ ● ] Three protocols in progress

Sonari Labs. Sonari 研究室.

Some of what we have built does not belong inside a product.
It belongs in a paper, a specification, a benchmark. 我们已经构建的一部分,不属于任何一个产品。
它属于一份 论文、一份 规范、一份 基准。

Sonari Labs is the research arm of Sonari — converting seven years of production knowledge into citable artifacts at a deliberate cadence. Founded by an IEEE-published voice AI researcher with 9 invention patents (incl. US 8160238), Labs publishes the protocols, benchmarks, and specifications that Sonari Platform and Sonari Scholar implement. Published in public, indexed where the field can find it, and made citable so the next generation of stateful AI can be built on top of it. Sonari Labs 是 Sonari 的研究部门 —— 按照固定节奏,把七年的生产知识转化为可被引用的学术产出。由具有 IEEE 论文记录的语音 AI 研究者创立,持有 9 项发明专利(包括 US 8160238)。Labs 负责发布 Sonari Platform 和 Sonari Scholar 共同实现的协议、基准和规范。公开发表、可被领域检索、成为可被引用的资产 —— 让下一代 stateful AI 可以建立在它之上。

See the three protocols → 查看三项协议 → GitHub · sonari-labs →
Founder & Director创始人 / 主任
Ronnie Cheng · IEEE
Patents专利
9 · US 8160238 + PCT
Production data生产数据自
2019 · 4 markets
Digital Self liveDigital Self 上线
Nov 2024
07 / WHY A LAB

We have been doing research
without naming it. 我们一直在做研究 ——
只是从未为之命名。

For seven years, four products, and four emerging markets, Sonari has been running what — in retrospect — looks much more like an applied research program than a consumer software business. The data we sit on is not replicable inside any existing affective computing lab. The protocol-level decisions made when we shipped Digital Self in November 2024 are still ahead of any published reference in the field.

This was tolerable when the work was new. It is not tolerable now. If we do not name what we have built, someone with less data will name it first — and the next decade of stateful AI will be built on top of their words, not ours.

Sonari Labs is the part of the company that publishes. It exists to convert production knowledge into citable artifacts — preprints, benchmarks, open specifications — at a deliberate cadence. It does not pivot the company. Every protocol Sonari Labs is working on corresponds to something that already ships in our products. The lab is the act of making the existing visible.

七年时间、四款产品、四个新兴市场 —— 回头看,Sonari 一直运行的不像一个消费软件公司,更像一个应用研究项目。我们手上的数据,任何现有的情感计算实验室都无法复制。2024 年 11 月 Digital Self 上线时所作的协议级决定,至今领先于学术界已发表的任何参考。

这一切在产品早期是可以容忍的。今天不行了。如果我们不为我们构建的东西命名,数据更少的人会抢先命名 —— 那么下一个十年的 stateful AI,会建立在他们的术语之上,不是我们的。

Sonari Labs 是公司里负责"对外发表"的部分。它存在的意义,是把生产环境里的知识,按固定节奏转化为可被引用的学术产出 —— 预印本、基准、开放规范。它不让公司转向。Sonari Labs 正在做的每一个协议,都对应公司产品里已经在跑的东西。Labs 这件事,就是把已有的、变成可被看见的。

Eighteen months of shipping a thing is not the same as eighteen months of thinking about shipping a thing. It is the only lead in this category that cannot be bought — but it is also the only lead that disappears if no one ever writes it down. 把一件事 ship 十八个月,和思考十八个月该不该 ship,不是同一件事。这是这个赛道里唯一无法被钱买到的领先 —— 但也是唯一会"如果你不写下来就消失"的领先。 — Ronnie Cheng · April 2026
08 / LINEAGE

The credential is real.
It has just been offstage. 资历是真的 ——
只是一直在幕后。

From research
to operator
and back. 从研究到产品
再回到研究。

Most consumer AI founders entered the category from product. Ronnie entered from voice algorithm research at Huawei Mobile, then spent fifteen years applying it at consumer scale across four markets. The published record is below. 大多数消费 AI 创始人从产品端进入这个赛道。Ronnie 是从华为移动的语音算法研究端进入的,然后花了十五年在四个市场把研究做成了消费规模的产品。已发表的记录如下。

  • ~2007
    US Patent 8160238 · double-talk detection — foundational for modern AEC systems. Filed during tenure as Head of Voice Algorithm at Huawei Mobile. Patent record 美国专利 US 8160238 · 双向通话检测 —— 现代回声消除系统的基础技术。在华为移动语音算法部门负责人任内提交。 专利原件
  • ~2008
    PCT international patent + 7 invention patents · nine total — the IP foundation Sonari voice infrastructure descends from. PCT 国际专利 + 7 项发明专利 · 共九项 —— Sonari 语音基础设施技术血统的源头。
  • 2006
    IEEE publication · speech processing. IEEE record IEEE 论文发表 · 语音处理领域。 IEEE 原文
  • 2011–2024
    Fifteen years of voice-first product operation across MaiChang, KongEr, Sigo, peplive, pepstar, veco, uho — four emerging markets. The research that did not have a name yet. 十五年的语音优先产品运营 —— 跨 MaiChang(麦唱)、KongEr(空耳)、Sigo、peplive、pepstar、veco、uho,四个新兴市场。当时还没有名字的研究。
  • Nov 2024
    Production deployment of Digital Self in peplive — voice cloning, persona persistence, autonomous social tasks, sovereign user ownership. One of the earliest deployments of this configuration in a voice-social context anywhere in the world. Digital Self 在 peplive 中投入生产 —— 语音克隆、人格持久化、自主社交任务、用户主权。这一配置在 voice-social 场景下,是全球最早的生产部署之一。
  • Apr 2026
    Sonari Labs established · publishing begins. First arXiv preprint targeted Q3 2026. Three protocols in active development. Sonari Labs 正式成立 · 公开发表开始。首篇 arXiv 预印本目标 Q3 2026。三项协议正在推进中。
09 / PROTOCOLS IN PROGRESS

Three protocols.
One thesis underneath them. 三项协议。
其下,同一个 thesis。

The thesis is that stateful, relational, continuation-grade AI is a distinct category from the AI companion paradigm — and that the category needs its own protocols, benchmarks, and standards before incumbents from the closed-loop companion era can claim them by default. 这个 thesis 是:stateful、relational、continuation-grade 的 AI,是一个区别于"AI 伴侣"范式的独立赛道 —— 而这个赛道需要属于自己的协议、基准和标准,在 companion 时代的旧巨头默认占领之前。

— Protocol · 01 Position paper · drafting 立场论文 · 撰写中

ACP Affective Continuity Protocol

— Voice · Affect · Longitudinal — 语音 · 情感 · 长期纵向

Affective computing has solved the single-shot problem: given a voice clip, identify the emotion. The unsolved problem is the longitudinal one — how affective state persists, drifts, and remains faithful to its principal across days, weeks, and shifting relational contexts. Every existing benchmark (IEMOCAP, EMOTIC, AffectNet, MELD) is session-bounded by construction. ACP defines the missing protocol — and Sonari is uniquely positioned to do so, because the longitudinal voice corpus required does not exist outside of our production deployments. 情感计算已经解决了单次任务:给一段语音,识别情绪。未解决的是长期问题 —— 情感状态如何在天、周、月的尺度上持续、漂移,并在不断变化的关系语境中对其本体保持忠实。所有现有基准(IEMOCAP / EMOTIC / AffectNet / MELD)在构造上都是单 session 边界的。ACP 定义这个缺失的协议 —— Sonari 处于独特位置来做这件事,因为所需的长期纵向语音语料库,在我们的生产部署之外并不存在。

Target venue目标会议 IEEE T-AFFC · Interspeech · ICASSP Bench AffectiveContinuity-Bench v0.1 SDK acp-sdk · MIT
— Protocol · 02 Position paper · drafting 立场论文 · 撰写中

SPVP Synthetic Participant Validation

— LLM-as-Participant · Social Science · Validity — LLM 作为受试者 · 社会科学 · 效度

Princeton, Stanford, NYU have published prominent papers using GPT-4 to replicate human survey and behavioral experiments. The validity question is wide open. There is no validated standard for when LLM-generated participant data is admissible as social-science evidence. Sonari's Agent Society — a 24/7 in-production multi-agent simulation operating at scales orders of magnitude beyond academic demos — is the natural validation substrate. SPVP, developed jointly with Sonari Scholar, defines the admissibility protocol the field needs. 普林斯顿、斯坦福、NYU 已经发表了多篇用 GPT-4 替代人类做问卷和行为实验的高引用论文。效度问题完全没有结论。没有任何标准能说明 LLM 生成的受试者数据,什么时候能作为社会科学证据被接受。Sonari 的 Agent Society —— 一个 7×24 在生产环境运行的多智能体模拟系统,规模比学术 demo 高几个量级 —— 是天然的验证基底。SPVP 与 Sonari Scholar 联合开发,定义这个领域亟需的可采纳性协议。

Target venue目标期刊 Nature Human Behaviour · PNAS · ACL Bench SocialSim-Bench · EN/ZH bilingual Joint work联合工作 Sonari Labs × Sonari Scholar
— Protocol · 03 Preprint · Q3 2026 预印本 · Q3 2026

RMSpec Relational Memory Specification

— Memory · Relation · Continuation — 记忆 · 关系 · 延续

All current AI memory frameworks — Mem0, Letta, Zep, native LLM memory — treat memory as user-attribute storage. None model the relationship itself between a user and an AI, between an AI and third parties, across time. Sonari's L5 Relational Memory has been in production since 2024 and is the only known system in this category at consumer scale. RMSpec is the open specification that formalizes the model and invites adoption — the connective tissue between Sonari Platform (voice-social) and Sonari Scholar (academic) that proves the spec generalizes across two materially different domains. 目前所有 AI 记忆框架 —— Mem0、Letta、Zep、各家 LLM 原生记忆 —— 都把记忆当作用户属性的存储。它们都不建模关系本身:用户与 AI 之间、AI 与第三方之间、跨时间的关系。Sonari 的 L5 Relational Memory 自 2024 年起在生产环境运行,是这一类系统中已知唯一的消费规模实现。RMSpec 是开放规范,把这一模型形式化并邀请采用 —— 它是 Sonari Platform(voice-social)与 Sonari Scholar(学术)之间的连接组织,证明这一规范在两个本质不同的领域都能成立。

Target venue目标会议 NeurIPS · ICML · FAccT Spec github.com/sonari-labs/rmspec · MIT First preprint首篇预印本 arXiv · Q3 2026
10 / WHAT WE PUBLISH · CADENCE

At least one public artifact every quarter. 每个季度,至少一份公开产出。

Quarters that fail this test are diagnosed as a structural failure of the lab — not a delay. The protocol matters. The schedule is how we know the protocol matters. 没能完成这个测试的季度,会被诊断为 lab 的结构性失败 —— 不是"延期"。协议重要。这个时间表,是我们让协议持续重要的方式。

Q2 · 2026
RMSpec v0.1 on GitHub · ACP position paper drafted · Sonari Labs page live RMSpec v0.1 上 GitHub · ACP 立场论文初稿 · Sonari Labs 页面上线
Q3 · 2026
First arXiv preprint (RM-Voice) · ACP position on arXiv · AffectiveContinuity-Bench v0.1 首篇 arXiv 预印本 (RM-Voice) · ACP 立场论文上 arXiv · AffectiveContinuity-Bench v0.1
Q4 · 2026
SPVP position paper · SocialSim-Bench v0.1 · first conference workshop submission SPVP 立场论文 · SocialSim-Bench v0.1 · 首篇 workshop 投稿
Q1 · 2027
First full conference submission (ICASSP / Interspeech / NeurIPS) · Sonari Scholar RM-Academic preprint 首篇正式会议投稿 (ICASSP / Interspeech / NeurIPS) · Sonari Scholar RM-Academic 预印本
11 / COLLABORATE

We are looking for specific people. 我们在寻找特定的人。

Sonari Labs runs lean by design — full-time research is bounded, but collaboration is not. Each of the three protocols benefits from one or two external co-authors with deep domain expertise. CV credit is real; intellectual partnership is the actual point. Sonari Labs 故意保持小而精 —— 全职研究的规模有限,但合作没有上限。三项协议每一项都受益于一两位有深度领域专长的外部合作者。CV credit 是实在的;真正的目的是智识上的伙伴关系。

For ACP
Affective computing PhDs · voice AI researchers · longitudinal study designers. If your prior work touches IEMOCAP / EMOTIC / continuous affect — write. 情感计算方向 PhD · 语音 AI 研究者 · 长期纵向研究设计者。如果你的过往工作涉及 IEMOCAP / EMOTIC / 连续情感 —— 写信来。
Email Ronnie 联系 Ronnie
For SPVP
Computational social scientists · LLM-as-participant skeptics · methodologists. If you have published on synthetic data validity, we want to talk. 计算社科研究者 · 对 LLM-as-participant 持怀疑态度的人 · 方法论学者。如果你发表过关于合成数据效度的研究,我们想聊聊。
Email Ronnie 联系 Ronnie
For RMSpec
Memory infrastructure builders · agent runtime researchers · open-spec authors. The spec is open. Implementations welcome. 记忆基础设施开发者 · 智能体 runtime 研究者 · 开放规范作者。规范是开放的。欢迎实现。
Email Ronnie 联系 Ronnie
Press · Investors 媒体 · 投资人
Series A · 2026. Sonari is fundraising under the applied AI lab + consumer products framing. Series A · 2026。Sonari 以"应用 AI lab + 消费产品"的框架对外融资。
Email Ronnie 联系 Ronnie

Some of what we have built belongs in a paper.
Some belongs in a specification.
Some belongs in a benchmark.
All of it has been in production for a while. 我们构建的有些东西,属于一份 论文。
有些,属于一份 规范。
有些,属于一份 基准。
它们都已经在生产环境里跑了很久。

We are simply choosing, at last, to write it down. 我们只是,终于决定,把它们写下来。

Email ronnie@sonari.io → 写信给 ronnie@sonari.io → GitHub · sonari-labs →