There is no Doubao IME on Linux, so Jackson wired Doubao's streaming speech recognition into Omarchy: hold F9 and speak, live text shows at the bottom of the screen, and about a second after release it lands at the cursor; Claude Code can install it from AGENTS.md
Developer Jackson (@Jacksonlin945) uses DHH's Omarchy (Arch + Hyprland) as his daily desktop, but ByteDance's Doubao input method, which many Chinese users consider the best Chinese dictation, has no Linux build. Instead of the IME he calls Volcengine's public Doubao streaming ASR API from a small resident Python service, doubao-dictate: hold F9 and audio streams while you talk, a floating bar shows live captions, and on release the server does a second, non-streaming pass and the text is pasted at the cursor. v0.2 is open source (MIT), and the README and AGENTS.md describe letting Claude Code or Codex do the install. He reports that for a 17-second sentence the wait after release dropped from 2.5s+ in v0.1 to 0.7–1s. We don't have an Omarchy machine or Volcengine credentials, so we only read the code and syntax-checked it.
Visit site 豆包流式语音识别(火山引擎)+ omarchy-doubao ↗- Pricing
- Paid
- Region
- China
- Where it's from
- Chinese community
- Added
- Oct 24, 2026
- Version we used
- omarchy-doubao v0.2.0, main @ 5cf4ca1 (2026-10-22); Doubao streaming ASR model 2.0 (volc.seedasr.sauc.duration)
- Categories
- ProductivityAudio & Music
- Original author
- Jackson|坚叔 @Jacksonlin945 ↗
- Our test
- Not tested by us
- Cost / time
- The code is free (MIT). Recognition uses Volcengine's Doubao streaming ASR 2.0, billed by audio duration; the author says personal dictation volume is small, and current pricing is in the console. You need a Volcengine account, which is a China-side cloud account with real-name verification.
The task
Get Doubao-quality Chinese dictation on a Linux desktop (Omarchy / Hyprland): mixed Chinese-English, automatic punctuation, live text while speaking, working in terminals, browsers and the Claude Code prompt.
Steps
- Core idea: you don't need the Doubao IME, only the recognition engine behind it. Volcengine exposes the Doubao streaming ASR model as an API, so nothing runs locally.
- Get credentials: in the Volcengine "Doubao Speech" console, enable "Doubao streaming ASR model 2.0" with per-duration billing; new consoles give you one API Key, older accounts use App ID + Access Token.
- Match the resource ID: model 2.0 per-duration is volc.seedasr.sauc.duration (default), 1.0 is volc.bigasr.sauc.duration, and concurrent plans swap .duration for .concurrent. A wrong prefix fails at the handshake.
- Install: git clone https://github.com/jacksonlin-infra/omarchy-doubao.git, then ./install.sh --api-key '<your API key>' (older consoles: --app-id + --access-token). The script installs missing packages (wtype, wl-clipboard, python-websockets, gtk4-layer-shell, etc.), writes credentials to ~/.config/doubao-dictate/doubao.env (mode 600), disables Omarchy's built-in Voxtype service (without removing it), appends a marked F9 / Super+Ctrl+X block to ~/.config/hypr/bindings.lua, starts the user service and runs ./check.sh.
- Or hand it to a coding agent: send the sentence in the Prompt box plus the repo link to Claude Code / Codex on that machine. It follows AGENTS.md and only asks you for credentials and, if needed, to type a sudo password in your own terminal. AGENTS.md tells the agent not to invent, echo or commit credentials. (The prompt is in Chinese as the author wrote it; in English: "Following the AGENTS.md at https://github.com/jacksonlin-infra/omarchy-doubao, install and verify Doubao voice input on this Omarchy machine.")
- Use it: hold F9 to talk, release to insert; Super+Ctrl+X toggles for long passages (max 300 seconds per take). Watch the live captions in the bottom bar; the second pass after release fixes most of their mistakes.
- Check and debug: ./check.sh does a real Doubao auth round-trip; ./check.sh clip.wav runs a 16 kHz mono WAV through the full pipeline; set DOUBAO_DEBUG=1 in doubao.env and follow journalctl --user -u doubao-dictate -f for per-step timestamps.
- Remove: ./uninstall.sh removes the service and the key-binding block and re-enables Voxtype; --purge also deletes the credentials.
Prompt
Following the AGENTS.md at https://github.com/jacksonlin-infra/omarchy-doubao, install and verify Doubao voice input on this Omarchy machine.
What the author says
He says v0.1 (Voxtype plus a protocol shim, built in about half an hour by his account) had to upload the whole clip after you stopped, so text took 2.5s+ after release. v0.2 switches to the bigmodel_async bidirectional streaming endpoint with a second recognition pass: live captions while speaking, and for a 17-second sentence the final text arrives 0.7–1s after release, as accurate as whole-clip recognition. It works in browsers, terminals, editors and Claude Code / Codex prompts. The README's demo GIF notes the speech was TTS fed through a virtual microphone. He says he now dictates most docs, replies and AI prompts.
Our test
We have no Omarchy / Hyprland desktop, microphone or Volcengine credentials, so we didn't run any dictation; latency figures are the author's. We only cloned the repo (main @ 5cf4ca1) on our test machine and read it: bin/doubao-dictate (about 930 lines, Python stdlib plus websockets/GTK) passes py_compile, and install.sh / check.sh / uninstall.sh pass bash -n. The code matches the write-up: the default endpoint is bigmodel_async, enable_nonstream is only turned on for that endpoint, audio goes out in 200 ms packets, and insertion uses wl-copy plus hyprctl sendshortcut, restoring the clipboard afterwards.
How to get in
The write-up is a Chinese X Article (login required); the GitHub README and AGENTS.md are in Chinese, but the commands read fine as-is and an agent can follow AGENTS.md. You need a Volcengine (ByteDance's China cloud) account, which typically requires real-name verification; the repo is configured for Volcengine's openspeech.bytedance.com endpoint, not the international BytePlus console.
Who it suits
People on Omarchy or another Arch + Hyprland desktop who want Chinese or mixed Chinese-English dictation, especially developers who dictate prompts to Claude Code / Codex. Also a reference pattern for adding dictation to a platform with no official client by calling a cloud ASR API.
Pitfalls
- The resource ID must match the model version and billing type you enabled; a wrong prefix fails at the handshake (403 requested resource not granted).
- Only one of the three streaming endpoints works: .../sauc/bigmodel returns 400 resourceId ... is not allowed for model 2.0 per-duration resources, bigmodel_nostream only returns after you finish (no live text), and only bigmodel_async supports the second recognition pass.
- Don't type Chinese character by character with wtype: Omarchy runs fcitx5 by default, and some characters' keycodes get read as Backspace or Tab, so characters vanish and opencode even switched itself from Plan to Build. He switched to pasting via the clipboard (Shift+Insert in terminals, Ctrl+V elsewhere) and restoring the previous clipboard text afterwards; an image on the clipboard is not restored.
- Filler words like 嗯 / 这个 / 那个 are kept; he found Doubao's DDC (disfluency removal) switch doesn't remove them and plans an LLM clean-up pass next.
- Upgrading from v0.1 isn't in place: uninstall with v0.1's own script first (git show v0.1.0:uninstall.sh | bash -s -- --purge) and re-enter credentials.
- Audio is sent to Volcengine for recognition (only while the hotkey is held). Outside China, latency depends on your route to openspeech.bytedance.com; if you use a proxy, he suggests sending that domain direct.
- Writing key bindings on Omarchy 3 (bindings.conf) is untested by the author; other Hyprland setups need three bind lines added by hand. The early v0.1 article had a placeholder link for the shim script; v0.2 puts the full code in the repo.
More in this category
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Have fresh-context AI reviewers poke holes before you codeCodingEGMCodexClaude CodePlugin free; reviews use your model quota◐ Partly testedOverseas
- 1,050 memory recalls audited: what makes agent memory pay offCodingNowledge MemCodexClaude CodeFree tier: 50 memories; review uses model quota○ Not tested by usChina
- Isolate an always-on agent on its own machine, then red-team itAutomation & AgentsOpenClawClaude CodeOllama$20/mo Ollama Cloud + a Claude plan○ Not tested by usOverseas
- A to-do list beside your agents, one AI chat per taskProductivityToday BoardcmuxClaude CodeFree, open source; needs macOS + cmux◐ Partly testedChina