Audio Inbox

🎤 Audio Inbox

One-click voice recording → speech-to-text → AI summary, saved as Markdown inside your vault. Desktop & mobile · Configurable speech recognition and AI services · Bring your own API credentials

⭐ If this plugin helps you, a Star is appreciated!

---

✨ Features

• 🎙️ One-click recording — tap the mic, speak, stop. No extra steps.
• ⏱️ Long recordings — record up to 60 minutes by default (configurable), compressed as WebM/Opus so hour-long audio stays small.
• 🧠 Smart transcription — SiliconFlow, compatible transcription APIs, or Doubao Voice standard 2.0 / flash (SiliconFlow default TeleAI/TeleSpeechASR; switch to the free FunAudioLLM/SenseVoiceSmall tier if you prefer).
• 🤖 AI summarization — DeepSeek (default deepseek-v4-flash) generates a structured note with a summary and an extracted to-do list.
• 💭 Memo & Reminder modes — the AI classifies each recording as a reminder, a memo, or both, and corrects speech-recognition errors automatically.
• 📱 Mobile support — a draggable floating record button on iOS/Android, which you can hide in settings.
• 📂 Organized output — every recording becomes its own file, filed by month and day (see below).
• 🍎 Apple Reminders sync — a clean to-do list is copied to your clipboard for iOS Shortcuts.
• 🌐 China-friendly — no Google APIs, no VPN. Both default services are domestic.
• 💾 Audio kept by default — original recordings are preserved after processing (can be auto-deleted in settings).

🆚 Why Audio Inbox?

---

🔄 How it works

🎤 Record  →  🎧 Transcribe (STT)  →  🤖 AI summary  →  📝 Markdown note

Each recording is saved as an independent file:

VoiceNotes/
  2026-09/                 ← month
    2026-09-23/            ← day
      备忘-练车准备.md      ← memo  (one file per recording)
      待办-买书.md          ← todo

• File names use a short AI-generated title (e.g. 备忘-练车准备.md, 待办-买书.md).
• Same-day, same-title collisions get a time suffix (备忘-练车-1840.md).

---

🚀 Quick start

1. Get two API keys (both China-friendly, no VPN)

1. Speech-recognition credentials — SiliconFlow, a compatible service, or Doubao Voice (see the requirements below). Availability, trial quotas and fees depend on the provider and your account.
2. DeepSeek API key — for AI summarization → https://platform.deepseek.com/

2. Install

Option A — Community plugins (recommended) Obsidian → Settings → Community plugins → Browse → search "Audio Inbox" → Install → Enable.

Option B — Manual Download main.js, manifest.json, and styles.css from the latest Release, place them in <vault>/.obsidian/plugins/audio-inbox/, then enable the plugin in Settings → Community plugins.

3. Configure

Open Settings → Audio Inbox, select the speech-recognition service, and enter its credentials. Configure the AI summarization service separately. The compatible-service defaults are:

• STT model: TeleAI/TeleSpeechASR (or FunAudioLLM/SenseVoiceSmall for the free tier)
• Summary model: deepseek-v4-flash

4. Record

• Desktop: click the mic icon in the left ribbon
• Mobile: tap the floating button (drag to reposition)
• Command palette: Ctrl/Cmd + P → "开始语音笔记（录音）"

⏱️ For best accuracy, keep recordings reasonable in length. Long recordings are supported but consume more API quota.

---

豆包语音：开通条件与配置要求

豆包在本插件中负责语音转文字（ASR）。AI 总结仍使用你单独配置的、兼容 Chat Completions 的服务，需另外填写该服务的地址、模型和 Key。

使用前需满足：

• 火山引擎账号已开通所选的录音文件识别服务，并有可用额度或余额。服务商可能收费；插件免费不等于豆包 API 免费，实际计费、试用额度和开通要求以控制台为准。
• 凭证与服务属于同一个项目；旧版控制台须属于同一个应用。新版使用豆包语音 API Key，旧版使用 APP ID + Access Token，两种鉴权方式择一。
• 豆包聊天会员、火山方舟聊天模型 Key、流式语音识别、语音合成或声音复刻权限，均不能替代上述录音文件识别权限。
• Obsidian 能联网访问语音服务；直接录音需授予麦克风权限。手机录音时保持应用在前台，锁屏或切到后台可能中断。

配置步骤：

1. 打开火山引擎「豆包语音」控制台，选择已开通服务的项目，获取对应语音凭证。
2. Obsidian → 设置 → Audio Inbox → 语音识别服务，选择 豆包语音。
3. 将「豆包识别模式」设为 标准版 2.0（或已获授权的极速版）。
4. 新版控制台选择「API Key」，仅粘贴 Key 值；旧版选择「APP ID + Access Token」，分别粘贴两项值。**不要粘贴整段 curl、引号、Header 名称或示例代码。**接口地址、资源 ID 和模型由插件按模式自动设置。
5. 保持 AI 总结配置完整，先录制 5～10 秒测试语音，再处理较长录音。

文件、等待与数据说明：

• 支持 WAV、MP3、OGG Opus；WebM/MP4 等录音会尝试转换为 16 kHz 单声道 WAV。源文件及转换后的文件均不得超过 100 MB。转换会增加内存占用与文件大小，长录音建议分段；60 分钟的录音设置不代表每家接口都能处理 60 分钟。
• 标准版提交任务后，每 2 秒查询一次，最多查询 150 次，并在约 5 分钟后停止继续查询（单次网络请求耗时可能使实际等待更长）。超时或失败保留原音频。重新处理会提交新任务，可能再次计费；插件不会自动恢复先前任务。
• 录音直接发送至火山引擎语音接口，无需网盘公开链接；识别文字随后发送至你配置的 AI 总结服务。凭证保存在知识库的插件配置文件中；如同步知识库，请自行管理该配置文件的同步范围。不要在公开截图或 issue 中包含凭证。
• 新配置默认推荐标准版；此前测试版已选择豆包的配置保留极速版选择，可手动切换。原兼容接口和豆包凭证分别保存。
• 标准版 Base64 上传已通过真实中文合成语音测试。官方标准版文档主要列出 URL 输入，Base64 属于实测可用行为；不承诺服务端未来保持此未明确文档化的输入方式。极速版及旧版凭证已通过模拟测试，尚未完成本次真实账号验证。

错误排查： 401 / Invalid X-Api-Key：检查语音凭证和鉴权方式；403 / requested resource not granted：检查同一项目是否开通所选资源。仅开通标准版时切换到标准版 2.0，无需继续寻找极速版。额度不足请检查账单；排队超时先等待再考虑重试。报告问题时可提供错误码和日志 ID，隐藏凭证及私密转写内容。

官方参考：标准版 API、极速版 API、API Key 使用。

⚙️ Settings

---

🍎 iOS Shortcuts (Apple Reminders sync)

After each recording, the plugin copies a clean to-do list to the clipboard. Create a Shortcut to import them:

---

❓ Troubleshooting

• Nothing recognized / empty note — check that your API key is correct and that the recording contains speech. Open the developer console (Ctrl+Shift+I) to see logs.
• "Model Not Exist" — make sure the summary model is deepseek-v4-flash (DeepSeek retired deepseek-chat / deepseek-reasoner on 2026-07-24).
• Balance / quota errors — top up your SiliconFlow or DeepSeek account.
• Microphone unavailable — grant Obsidian microphone permission, or record with Obsidian's built-in recorder and run "处理录音文件夹" (Process recording folder) from the command palette.

---

🔧 Development

git clone https://github.com/andsea007/obsidian-audio-inbox.git
cd obsidian-audio-inbox
npm install
npm run dev      # watch mode
npm run build    # production build → main.js

---

📄 License

MIT