Qiaomu Audio

Read any Obsidian note aloud with one click, and dictate speech straight into your notes.

Official directory · 中文说明 · Releases · Help and feedback

Read aloud

• Use the speaker button in any Markdown note. Read selected text, start at the cursor, or start from the visible text in reading view.
• Pause, resume, move between sentences, change speed and choose a voice in a compact playback bar. The current sentence is highlighted in the editor; reading view follows the relevant text. Manual scrolling pauses page following until you choose to return.
• Frontmatter, code blocks, formulas, comments, embedded media and raw URLs are skipped. Links read their visible text.
• Choose system speech, Mac speech, Edge TTS, offline Kokoro/MeloTTS, OpenAI and compatible services, Qwen, Doubao, MiniMax, SiliconFlow, ElevenLabs or Azure.
• Import saved reading settings and keys from Qiaomu Reader. Matching downloaded model data is reused.

Dictate

Use the microphone button or the dictation command. Finish to insert recognised text at the captured cursor position; Escape cancels. Failed recognition keeps the recording available for retry or saving.

• Local SenseVoice supports Chinese, English, Japanese, Korean and Cantonese on desktop. An external runtime setup is required once; then download about 156 MB of model data in settings.
• Cloud recognition supports SiliconFlow, Doubao, GLM, StepFun, Xiaomi MiMo, Groq and OpenAI, plus local/custom compatible services. Save multiple configurations and switch between them.
• Choose recognition after finishing, or recognition of each phrase after a pause. Optional saved audio uses the vault's attachment folder.

Cloud setup

Choose a provider, follow Get a key or Setup guide, paste its API key, and test the voice or connection. Required fields such as Azure region and custom service URL are visible. Service addresses for named providers and low-frequency options are under Advanced.

Keys are hidden by default, can be shown temporarily, and are stored in Obsidian SecretStorage rather than plugin data.json. Speech-to-text keys are isolated by provider, and existing keys are migrated to the saved provider. Switching settings or cancelling a test clears its result; a saved configuration is never labelled connected until its test succeeds.

Voice tests send a short sample sentence. Recognition tests send one second of generated audio without opening the microphone. Tests may incur a small provider charge. Tests run only when requested.

Privacy, network use and external files

• System speech, Mac speech and local model inference do not upload your text or recordings. System voices may depend on your device's OS configuration.
• Edge TTS sends reading text to Microsoft's speech service. It requires network access, uses an unofficial interoperability protocol, needs no API key, and is desktop-only. Availability is not guaranteed; system speech is an alternative.
• Cloud providers receive only the text or recording required for the selected action. Their accounts, service access and usage charges are governed by their terms. Local compatible services use localhost; custom remote services receive the same request data.
• Local model data is downloaded only after you choose Download and use, from fixed upstream Sherpa GitHub releases. File sizes and SHA-256 checksums are verified. Model data, worker configuration, temporary audio and runtime files are outside the vault, in platform application data/cache directories, so multiple vaults can share model data without syncing it. macOS uses ~/Library/Application Support/qiaomu-audio/speech; Windows uses %LOCALAPPDATA%\qiaomu-audio\speech; Linux uses $XDG_DATA_HOME/qiaomu-audio/speech or ~/.local/share/qiaomu-audio/speech.
• Python/Sherpa dependencies are prepared by a separate optional installer run outside Obsidian. The plugin does not install or update executable dependencies. See local runtime setup for the external tool's downloads and storage.
• There is no telemetry. Provider console/help links open in your browser without API credentials. About-page support images are requested from radio.qiaomu.ai only when the About section is opened.

Install

Open the official directory entry and choose Add to Obsidian, or search Qiaomu Audio in Community plugins once the client catalog includes it. Version 1.0.1 has completed its automated official review and its listing is public; the last recorded client-catalog check is in the release evidence.

For manual installation, use the latest release, put main.js, manifest.json and styles.css in <vault>/.obsidian/plugins/qiaomu-audio/, then enable Qiaomu Audio in Community plugins.

Mobile uses system speech and cloud services; Edge, Mac speech and managed local models are desktop features. The minimum Obsidian version is 1.11.4. Windows/Linux automated checks do not prove native device playback; mobile device playback and microphone acceptance have not yet been verified.

For other plugins

const audio = app.plugins.plugins['qiaomu-audio']?.api;
await audio?.speak('Hello');
const text = await audio?.transcribe(blob); // Blob / ArrayBuffer / Float32Array
// Start dictation in the current note:
audio?.startDictation();

Development

npm ci
npm run check

The check runs English/Chinese translation validation, ESLint, Node/Python tests and the production build. node scripts/e2e-local-asr.mjs downloads and checks the local recognition model when its external runtime is prepared. See architecture.

License and attribution

GPL-3.0-only. Reading code is adapted from Qiaomu Reader (GPL-3.0), and cloud recognition provider protocols from Qiaomu Clipper (MIT), both by 向阳乔木.

Edge TTS interoperability is based on rany2/edge-tts (LGPL-3.0); see the license. WebSocket transport uses ws (MIT; license included in the bundle).

Models are not distributed with the plugin: Kokoro (Apache-2.0), MeloTTS (MIT), SenseVoice Small (FunASR Model License). The optional Sherpa ONNX runtime is Apache-2.0 and is installed separately.