Linux voice typing features beyond dictation

Dictation turns speech into text. A fuller writing workflow also lets you change the cleanup strength, rewrite existing text, trigger recording from the mouse, and work on a draft before putting it into another app.

Bulbul brings those controls into one Linux app: Raw, Clean, and Polished cleanup; preset and custom AI transforms; click-to-toggle mouse mode; and a standalone scratchpad with transforms on selections. Snippets, a custom dictionary, history, and insights round out the workflow.

Those shared basics are not unique to Bulbul. OpenWhispr has snippets, a dictionary, and history; Handy has a dictionary and history too. OpenWhispr goes beyond dictation with a voice agent and notes. This guide compares the named Linux workflows, not every project or every feature on the market.

Bulbul has the fullest combination of the writing controls below. OpenWhispr's agent-based editing and notes are real alternatives, while Handy favors local speech. A different workflow is not the same as no feature.

OpenWhispr's voice agent can rewrite highlighted text after a spoken instruction such as 'make this more formal'. Its notes also support AI actions. Calling it dictation-only would hide features that matter when choosing a Linux writing app.

Handy's text-replacement proposal in PR #455 is still open in this check. Proposed code is not evidence of a shipped feature. This table also does not rule out external scripts, remappers, or custom post-processing prompts; those are different from built-in transform slots and mouse controls.

The comparison is based on official docs and repository source, not hands-on testing of every package. Check your installed release, since development source and packaged features can differ.

Bulbul transforms rewrite text you have already selected. The README lists Polish, Make Formal, Translate, Bulletize, and custom prompts, with Alt+1 through Alt+9 as the default transform bindings.

For example, select a rough paragraph and run a formal rewrite, or select a set of notes and turn it into bullets. These are selected-text actions, not a requirement to dictate a fresh instruction each time. Use the bindings in your installed Settings if you have changed them or a desktop shortcut conflicts.

Transforms and dictation cleanup solve different problems. Cleanup happens in the speech-to-text workflow; a transform operates on an existing selection. AI can change meaning or make mistakes, so review names, numbers, commands, and translated text before using the result.

OpenWhispr has an agent edit flow for selections too. The comparison is preset transform slots versus spoken agent instructions, not 'Bulbul can edit and OpenWhispr cannot'.

Bulbul offers three levels rather than a single cleanup switch. Raw fixes obvious errors, Clean removes fillers and fixes punctuation, and Polished rewrites for clarity. Clean is the documented default.

The README also documents a separate Polished hold-to-talk shortcut, Shift+Alt+P by default, which overrides the global cleanup level for that recording. Check your installed hotkeys before relying on it.

OpenWhispr describes cleanup as conservative and provides an on/off switch. Handy can toggle post-processing. Custom prompts may offer other behavior, but neither is documented here with Bulbul's Raw/Clean/Polished selector.

Mouse mode keeps the recording trigger under your hand. Click a supported middle, back, or forward button once to start recording, then click again to stop. The scratchpad gives you a separate place to write and apply transforms to selected text.

Mouse mode is click-to-toggle, not mouse-button hold-to-talk or voice control of the pointer. Linux input permissions matter: if virtual-mouse or exclusive access fails, the click can also reach the foreground app. Test in a harmless field before using a button that normally pastes or navigates Back or Forward.

Use the scratchpad when you want to check a draft and rewrite a selection before using it elsewhere. OpenWhispr also has notes with AI actions, so a scratchpad is not evidence that other Linux apps lack note-taking. Bulbul's standalone window and selection transforms are the specific workflow compared here.

No. Bulbul documents typing and hotkeys on both X11 and Wayland, but per-app styles are currently X11-only. Its README still lists Wayland app-detection for GNOME/wlroots as an open TODO.

Do not read 'Wayland supported' as a promise that app-aware tone selection works there. Choose your cleanup level and transforms explicitly rather than expecting Bulbul to detect every target app on Wayland.

Bulbul ships .deb, .rpm, and AppImage builds. Typing and hotkeys use evdev and a uinput virtual keyboard; .deb and .rpm installers configure access. An AppImage on Wayland can fall back to clipboard paste until you grant uinput access. Desktop environment, device permissions, and the installed version still matter.

OpenWhispr has x86-64 Linux builds with session-dependent input behavior. Handy documents limited Wayland support and may need typing tools and compositor shortcuts configured. These are setup differences, not reasons to pretend either app has no Linux version.

FluidVoice is macOS-only, so its Write Mode and per-app prompts are not Linux alternatives. Wispr Flow has no Linux version in this check. Their features on another operating system should not count as Linux features.

Yes, if it meets Bulbul's platform requirements and has a microphone, internet, and a Groq API key. Groq runs speech recognition, cleanup, and AI rewrites, so Bulbul does not require local speech-model downloads or a GPU for transcription.

The advantage is avoiding local model inference while keeping the writing controls. Your laptop still runs the interface, records audio, and uses the network. Connection quality, provider availability, and rate limits affect the experience; this is not a claim of identical latency on every cheap laptop or measured battery savings.

Bulbul talks directly to Groq with your key, without a Bulbul-owned transcription server between them. History, dictionary, snippets, and settings live in local SQLite. Their local storage does not mean text used for cleanup or transforms stays on-device.

Free app does not mean unlimited free cloud inference. Bulbul has no subscription or Bulbul-imposed word cap; Groq has a rate-limited free tier and may bill paid usage to your provider account. Speech, cleanup, and transforms can involve separate API requests.

OpenWhispr cloud mode can also avoid local speech inference. Handy and OpenWhispr local mode are better fits when offline speech matters. Keep cloud cleanup disabled if your rule is that audio and text must remain on-device. Bulbul's hosted pipeline needs internet.

Groq's API terms forbid model training on your inputs and outputs without your explicit permission or instruction. Audio still leaves your device for transcription, and text goes to Groq for cleanup and AI transforms.

Section 4.2 of the Groq Services Agreement says: "Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer."

Groq says inference content is not retained by default. Reliability troubleshooting or abuse investigations can temporarily retain inputs and outputs for up to 30 days, or longer when legally required. Usage metadata is always collected and does not include customer inputs or outputs.

Groq's data documentation says all customers may enable Zero Data Retention (ZDR) in Data Controls to prevent content retention for reliability and abuse monitoring. The agreement refers to eligible customers. Check your account controls and current terms; Bulbul does not automatically turn ZDR on. This is not a promise that no data of any kind is stored.

Bulbul's README also discloses anonymous usage statistics enabled by default, with an off-toggle in Settings > Privacy. It says those statistics exclude audio and transcripts. Local history, no-training terms, and ZDR controls are useful distinctions, but they do not turn hosted inference into offline processing.

Official product documentation, repository source, the open Handy proposal, and Groq API policies were checked on October 9, 2026. This is a comparison of the named apps, not an exhaustive market ranking or a hardware benchmark.