Release / 15 Oct 2026 / with v1.4.0, v1.5.0, v1.5.1, v1.5.2, v1.5.3, v1.6.0, v1.7.0, v1.8.0

Doubao Say v1.8.1

Dictate into macOS applications with the native menu-bar client.

All of Doubao Say

Before you upgrade to v1.8.1

2 to fix

New Linux installs default to `Ctrl+;`. GNOME changes unsupported saved evdev shortcuts, including Fn and modifier-only keys, to `Ctrl+;`; supported shortcuts and Disabled remain. Other desktops preserve saved shortcuts.

Behavior changedStart and finish dictation with tap or hold gestures using a chosen key or recorded shortcut.

Read the evidence

Approve the shortcut in GNOME's permission dialog and customize it there. Use tray/control-center Cancel; portal mode does not register global Esc. After denied authorization, restart to retry. Input delivery and Vibekey retain separate permissions.

Action requiredAffects: GNOME users upgrading from evdev triggers.

Read the evidence

quanru/doubao-say/blob/v1.4.0/README.md

During recording, another key (including Esc, Enter or a chord) finishes dictation and swallows that key and its release. During final polishing, a key inserts the original text. Use the menu's Cancel input to cancel; idle double-tap still sends Enter.

Behavior changedAffects: macOS users recording or waiting for final polishing.

Read the evidence

quanru/doubao-say/blob/v1.8.1/macos/README.zh-CN.md

Exact old default prompts migrate to contextual sound-error correction, explicit numbered lists and restrained paragraph breaks; custom prompts remain. Mac final polishing overlaps recognition but waits for final text and invalidates stale work.

Behavior changedPolish dictated text with editable Chinese and English prompts before inserting it.

Read the evidence

Install the new `PyJWT[crypto]>=2.10,<3` dependency. On Arch/Omarchy, rerun `install.sh` or `setup-omarchy.sh` to install `python-pyjwt` and `python-cryptography` before checking the runtime.

Action requiredAffects: Linux source/Git installations upgrading to ChatGPT-capable releases.

Read the evidence

What changed in v1.8.1

v1.4.0–v1.8.1 add GNOME portal shortcuts, macOS dictation, local vocabulary/history and speech evaluation. Updates can install on request; polishing supports ChatGPT accounts and revised defaults, with original-text fallback.

Also: GNOME-authorized triggers, personal vocabulary and corrections, 30-day history, speech comparisons, verified in-app updates and ChatGPT polishing.

v1.8.1: Linux app/plugin archives for Python 3.11–3.14; signed, notarized arm64/x86_64 DMGs for macOS 13+. Recognition needs an account/API key; Linux system libraries and desktop input permissions remain external.

New things you can do

Normal launch selects the portal; hold, tap and double-tap use authorized activation/release events. Denial never silently falls back to evdev.

AddedStart and stop dictation with a desktop-authorized GNOME shortcut.

Read the evidence

Tap/hold dictation, a menu-bar client and live overlay are available. New configs choose direct Accessibility insertion when supported, otherwise clipboard; failed input retains text for retry.

AddedDictate into macOS applications with Doubao, Volcengine or Deepgram.

Read the evidence

Settings save automatically; completed onboarding survives restart. Closing the last settings/sign-in window leaves menu-bar dictation running and hides Dock/Cmd+Tab entries; reopening restores them.

AddedConfigure macOS dictation and switch microphones from settings or the menu bar.

Read the evidence

History is enabled by default and preserves recognition, final text and status; newer entries include recognition, remaining polishing, delivery and total wait times.

AddedReview and reuse dictated text from a local rolling 30-day history.Inferred

Read the evidence

Vocabulary saves automatically; Volcengine hotwords prioritize confirmed spellings and show the actual subset. AI organization sends candidate terms/counts, not full history, and additions require review.

AddedSave personal terms and review suggestions to guide recognition and polishing.

Read the evidence

Select a wrong term in history, use Correct & remember, review the exact occurrence replacement and confirm. Original recognition and polishing output remain available.

AddedConfirm a transcript correction and reuse it as context for future polishing.

Read the evidence

The Mac/Linux evaluation page records samples and reports character errors, target accuracy, protected-term loss, failures and replay P50/P95. Drafts save automatically; evaluation cannot run alongside dictation or mic checks.

AddedCompare raw recognition, personal hotwords and polishing on the same recorded speech.

Read the evidence

`doubao-say-eval` or `python3 -m doubao_input.speech_eval` supports JSONL manifests, optional polishing and offline result scoring; reports are local and owner-only.

AddedReplay WAV recordings or score manually collected speech-recognition results from the command line.

Read the evidence

Account/model selection and usage controls preserve existing API settings. A quota limit pauses further plan requests; Manage usage and Retry plan usage support recovery without signing in again.

AddedPolish transcripts using an authorized ChatGPT account and an account-specific model.

Read the evidence

Fixed and improved

Updates now install on request with retained settings and rollback on installation failure. Checks try a public manifest then GitHub API; Mac shows progress, cancellation and specific errors, and verifies signing/notarization.

ImprovedCheck for a stable release and choose to download, verify, install and restart the application.

Read the evidence

With optional `wmctrl`, the X11 overlay requests above-window stacking without taking focus. Missing or failed stacking requests leave input working; fullscreen coverage remains window-manager dependent.

ImprovedPaste completed dictation into applications and terminals on native X11.

Read the evidence

quanru/doubao-say/blob/v1.5.0/README.md

API-key drafts survive onboarding/settings saves and provider changes; recording, microphone checks, evaluation, retry delivery and updates are coordinated to prevent overlapping operations.

Now worksSelect a PipeWire microphone and test dictation without inserting text into another application.

Read the evidence

quanru/doubao-say/releases/tag/v1.7.0

Seeded by @iluxav with Codex, from the release's public sources. Each change links to its evidence.