东南亚 AI 语音 Agent：2026 市场图谱
Three years ago, building a production voice agent required stitching together four or five vendors: a speech-to-text provider, an LLM, a text-to-speech engine, a telephony gateway, and some bespoke orchestration layer holding it all together. Today, Vapi and Retell have collapsed most of that stack into a single API surface.
The consolidation is real but fragile. Vapi’s core loop — microphone → Deepgram STT → GPT-4o → ElevenLabs TTS → speaker — works smoothly in a US English, low-latency demo environment. It starts fraying the moment you introduce Singlish code-switching, Bahasa Indonesia, or Tamil. The latency budget that feels acceptable at 600ms in San Francisco becomes conspicuous at 900ms in Jakarta, where mobile network variance is wider.
Southeast Asia’s enterprise voice landscape has three characteristics that western-optimised stacks consistently underestimate:
Multilingual codeswitching is the norm, not the edge case. A Singaporean customer service call will regularly toggle between English, Mandarin and Malay within a single utterance. No current voice agent handles this gracefully — they detect a language shift, pause, and either switch fully or hallucinate a mid-utterance translation.
Telephony is PSTN-heavy. SEA enterprises — banks, telcos, government agencies — overwhelmingly still route customer calls over PSTN/SIP, not VoIP-native WebRTC. Integrating a modern voice AI stack with a 20-year-old Avaya or Cisco CUCM deployment is not a solved problem.
Regulatory fragmentation is accelerating. Singapore’s PDPA amendments, Indonesia’s PDP Law (effective Oct 2024), and India’s DPDPA each impose different requirements on voice call recording, data residency, and consent logging. A voice agent that is compliant in SG may not be compliant in ID.
This is where the market gap opens. The global platforms — Vapi, Retell, Bland — are optimising for English-first enterprise workflows in North America and Western Europe. The SEA-specific opportunity is in:
Tencent Cloud’s TRTC + Chat stack has a structural advantage here: existing PSTN termination coverage across SEA, regional data centres in Singapore, Mumbai and Jakarta, and a pre-built compliance layer inherited from WeChat/Weixin governance requirements.
The 12-month trajectory I’d bet on: voice agent infrastructure commoditises faster than anyone expects (Twilio’s bet on AI is the signal), while the differentiation shifts to domain-specific fine-tuning and systems integration depth. The winner in SEA enterprise voice won’t be the platform with the best LLM — it will be the one that can reliably handle a codeswitched call in Singlish, log it compliantly under PDPA, and route it into a Salesforce workflow on a Cisco CUCM trunk.
That’s a harder problem than it sounds. It’s also a more defensible one.
三年前，要搭建一套生产级语音 Agent，你需要把四五个供应商拼凑在一起：语音转文字、大模型、文字转语音、电话网关，再加上某个把一切黏合起来的定制编排层。今天，Vapi 和 Retell 已经把这套拼图压缩进了一个 API 接口。
整合是真实发生的，但也是脆弱的。Vapi 的核心回路——麦克风 → Deepgram STT → GPT-4o → ElevenLabs TTS → 扬声器——在美式英语、低延迟的演示环境下运行流畅。但一旦引入新式英语（Singlish）语码混用、印度尼西亚语或泰米尔语，就开始出现裂缝。旧金山 600ms 感觉可接受的延迟，在雅加达——移动网络抖动更大的地方——900ms 就已经让人明显察觉。
东南亚的企业语音场景有三个特点，西方优化的技术栈一再低估：
多语言语码混用是常态，不是边缘案例。 一通新加坡客服电话，在单次对话中频繁切换英语、普通话和马来语是司空见惯的事。目前没有任何语音 Agent 能优雅地处理这一点——它们检测到语言切换时会停顿，然后要么完全切换语言，要么在切换过程中产生幻觉式翻译。
电话仍然以 PSTN 为主。 东南亚企业——银行、运营商、政府机构——绝大多数依然通过 PSTN/SIP 路由客户来电，而非原生 VoIP 的 WebRTC。把现代 AI 语音栈与一套 20 年历史的 Avaya 或 Cisco CUCM 系统对接，远不是一个已解决的问题。
监管碎片化正在加速。 新加坡 PDPA 修正案、印度尼西亚《个人数据保护法》（2024 年 10 月生效）、印度 DPDPA，在语音录音、数据驻留和知情同意记录方面各有不同要求。在新加坡合规的语音 Agent，在印尼未必合规。
这正是市场缺口打开的地方。全球平台——Vapi、Retell、Bland——正在为北美和西欧的英语优先企业工作流而优化。东南亚特有的机会在于：
腾讯云的 TRTC + Chat 在这里有结构性优势：覆盖东南亚的 PSTN 终端接入、新加坡、孟买和雅加达的区域数据中心，以及从微信/Weixin 治理要求继承而来的预置合规层。
我押注的 12 个月走势是：语音 Agent 基础设施的商品化速度会快于所有人的预期（Twilio 在 AI 上的押注就是信号），而差异化会转向垂直领域微调和系统集成深度。东南亚企业语音市场的赢家，不会是大模型最好的平台——而是能可靠处理新式英语语码混用通话、在 PDPA 框架下合规记录，并在 Cisco CUCM 中继上路由进 Salesforce 工作流的那一家。
这比听起来要难得多。也因此更难被复制。