Diction：iOS 语音键盘的自托管网关开源了

Apple 的原生听写有一个限制：它被沙箱隔离在当前 App 内，切 App 就失效。

Diction 绕开了这个限制。它注册成 iOS 系统键盘，作为自定义键盘挂载后，语音输入在任意文本框里都能用——备忘录、微信、邮件、终端，都没区别。

本次开源的是 Diction 的自托管后端网关。

GitHub: https://github.com/DictionLabs/Diction | MIT | Go

---

三种模式

模式	部署	隐私	费用
本地（on-device）	无服务器，纯离线	最高	免费
Diction One 云	官方托管	第三方处理	付费订阅
自托管网关	自备服务器	完全自控	基础设施自担

本地模式支持 99 种语言，完全离线，但依赖设备端模型精度。自托管模式精度取决于你挂的 STT 后端。

---

网关架构

网关是一个 Go 服务，协议是 WebSocket 流式传输，接口兼容 OpenAI Transcription API 规范。

这意味着任何实现了这套 API 的 STT 后端都可以挂上去。DictionLabs 官方发布了两个 Docker 镜像：

• dictionlabs/whisper-server — 基于 Whisper 的通用多语言方案
• dictionlabs/parakeet — NVIDIA Parakeet 模型，针对 25 种欧洲语言优化

安全层：键盘与网关之间端到端 AES-256-GCM 加密，录音不明文传输。

可选功能：LLM 后处理，对转录文本做润色或格式化。

---

自托管部署

docker compose up

启动后，在 iOS 的 Diction App 设置里填入网关地址，绑定完成。

---

诚实评估

没有 benchmark 支撑的数字：项目标榜”5x faster”，没有说比谁快、在什么条件下测的。这是未经证实的营销说法。

App Store 评分积累不足：产品还在非常早期，公开评价很少，无法从用户反馈判断真实体验。

Solo developer 风险：单人维护，长期支持和路线图不确定。

语言覆盖依赖后端选择：Parakeet 只覆盖 25 种欧洲语言，中文等亚洲语言需要用 Whisper 路径，精度和延迟自行评估。

无遥测是设计选择：对隐私有利，但也意味着没有公开的用量数据来验证项目活跃度。

---

适合谁

• 需要跨 App 语音输入、不信任云端 STT 的用户
• 想用私有服务器替代 Apple 沙箱的开发者
• 希望自控 STT 后端（精度、语言、成本）的场景

---

MIT 许可。DictionLabs 开发，网关代码 GitHub 已上线，iOS App 在 App Store（ID: 6759807364）。自报数字需打折核实。开源仅供学习参考。

---

Diction: Open-Source Self-Hosted Gateway for iOS Voice Dictation Keyboard

Apple’s native dictation has a hard constraint: it’s sandboxed per-app. Switch apps, it stops.

Diction works around this. It registers as an iOS system keyboard — once installed, voice input works in any text field: Notes, WhatsApp, email, terminal. No exceptions.

What just went open source is Diction’s self-hosted backend gateway.

GitHub: https://github.com/DictionLabs/Diction | MIT | Go

---

Three Modes

Mode	Deployment	Privacy	Cost
On-device	No server, fully offline	Maximum	Free
Diction One cloud	Official hosting	Third-party processes audio	Paid subscription
Self-hosted gateway	Your own server	Fully self-controlled	Infrastructure at your cost

On-device mode supports 99 languages, completely offline. Self-hosted accuracy depends on your chosen STT backend.

---

Gateway Architecture

The gateway is a Go service using WebSocket streaming, implementing the OpenAI Transcription API spec.

This means any STT backend that implements this spec can be plugged in. DictionLabs publishes two official Docker images:

• dictionlabs/whisper-server — general-purpose multilingual via Whisper
• dictionlabs/parakeet — NVIDIA Parakeet model, optimized for 25 European languages

Security layer: AES-256-GCM end-to-end encryption between the keyboard and gateway. Audio never travels in plaintext.

Optional: LLM post-processing pass on transcribed text for cleanup or formatting.

---

Self-Hosting

docker compose up

Then paste the gateway URL into Diction’s iOS app settings. Done.

---

Honest Assessment

Unverified claim: The project claims “5x faster” with no stated comparison baseline or benchmark methodology. Treat as marketing copy until validated.

Very early stage: App Store ratings insufficient to draw conclusions on real-world UX. Product Hunt listing shows no tracked reviews.

Solo developer risk: Single maintainer, uncertain roadmap longevity.

Language coverage: Parakeet covers only 25 European languages — CJK and other languages need the Whisper path; accuracy and latency should be tested for your specific use case.

No telemetry: Good for privacy, but also means no public usage data to validate adoption.

---

Who It’s For

• Users who need cross-app voice input without trusting cloud STT
• Developers who want to replace Apple’s sandbox with a private server
• Setups that require controlling STT backend (accuracy, language, cost)

---

MIT license. DictionLabs, gateway code on GitHub, iOS app on App Store (ID: 6759807364). Self-reported numbers should be independently verified. For technical reference only.