Voicebar

A voice-first Android keyboard that works alongside Gboard. Voicebar is a one-row bar under the app for talking, moving the cursor and editing; for typing, ABC pops up Voicebar's own full keyboard over the bar (hold it to hand the field to Gboard instead, and a button drawn over Gboard brings you back, already listening).

Install and set up

Download the APK from Releases and open it (allow your browser or file manager to install apps when Android asks). On first launch a setup wizard walks through everything Voicebar needs, one page each, checking each step as you come back from Android's screens:

1. Turn on the keyboards: Voicebar, and Gboard for typing (with a link to get Gboard if it's missing).
2. Switch to Voicebar in the keyboard picker.
3. Allow the microphone.
4. ElevenLabs key (speech to text): paste it and it's checked with ElevenLabs (a speech session opened with no audio) before it's saved, encrypted, on the phone.
5. Claude key (cleanup, optional): checked by listing models, then saved the same way.
6. The Gboard button (accessibility, optional): the way back from Gboard, hands-free mode and screen context. Android 13+ blocks accessibility for apps installed from a file; when Voicebar was installed that way, the page shows how to allow it (try once, then App info › ⋮ › Allow restricted settings).

Required steps have an “I'll do this later”; optional ones can be skipped. Run it again from Settings › How to use it › Run setup again. Updating an app that's already set up skips the wizard.

The bar

First-run tips: the first few times the bar opens, a purple bubble above it points at one control at a time: Voice, then (during your first recording) how to finish, then ABC, Cursor and ⌫ on later openings, and after a few dictations the “voice bar” commands. One per opening; using a feature retires its tip, “Got it” moves on, “Skip tips” stops them. Settings › How to use it has Show tips on the bar and Show all tips again.

• Seven equal buttons: ☰ · space · Cursor · ABC · Voice · ⌫ · ↵. ☰ is always the menu; the other three around Cursor, ABC and Voice are keys you choose, each from the same list (space, Backspace, Enter, Punctuation, paste, undo, select all, last dictation, fix last dictation). Hold one, or swipe up from it, to pick another on the spot (picking only sets the key; press it to use it), or choose in Settings ▸ Bar and Gboard. The two right of Voice stay put while you dictate; Enter there is a line break, never Send. As Backspace, a key deletes on a tap and keeps deleting on a tap then press-and-hold.
• Bar style (Settings ▸ Bar and Gboard, each shown as the real bar): Docked (default) across the bottom, or Floating: a small pill with a soft shadow, ⠿ · ☰ · Cursor · ABC · Voice · one key you choose (Backspace to start), each the docked keys' size. Drag the grip (⠿) to move it anywhere: it floats over the app, which keeps the whole screen; drop it near the bottom and it parks there again, with the app above it (where it starts). The app sits above the parked pill; the cursor pad opens above it at its width, and ABC opens the full-width keyboard (pull it down to get the pill back). From ☰, Collapse to the record pill leaves only ⠿, ☰ and Voice, floating over the app, which gets the whole screen and every touch outside the pill. Drag the pill anywhere; it stays on screen and comes back where you left it. Tap Voice and it grows into the recording pill in place (tap to finish, swipe right to send, left to cancel); hold it to fix by voice; ☰ ▸ Open the whole bar brings the pill back. ☰ ▸ Dock the bar / Floating bar switches style. ☰ ▸ Turn it upright stands the floating bar against a side edge: ⠿ · ☰ · Cursor · ABC · Voice · own key in a column of round keys, over the app (which keeps the whole screen); drag ⠿ anywhere and it stays where you drop it, except within about a fingertip (48 dp) of a side edge, where it snaps against that edge with a tick. Upright, Voice stays round while you dictate, with a small waveform pulsing in it: pull it up to send or down to cancel (it stretches to two keys tall; let go short of that and it springs back), tap to finish. Words still go into the box as you speak (with words only in the pill, nothing shows until you finish). The menu, own-key choices, the cursor pad and notes open beside the column; collapsed, it's ⠿ · ☰ · Voice. ☰ ▸ Turn it sideways lays it back down. Floating (open, collapsed or upright), a note or warning (a cleanup that didn't work, say) is a card of its own by the bar, in the bar's colors with an icon, up to two lines (three beside the upright column) so the whole message shows; the bar keeps its keys, tapping the card does what it offers (or puts it away), and Voice still starts listening.
• Voice: tap and talk. Voice becomes the recording pill (centered live waveform and a timer; until the speech service is connected, or Parakeet loaded, it says over the waveform: the mic is already recording and the audio is held, not lost); Cursor and ABC slide left into the places of ⋮ and the key after it, pushing those off the bar, in a quieter tone, and the two keys right of Voice stay where they were. Open the keyboard (ABC, or Gboard) within 4 s of listening starting and whatever was heard is thrown away, not cleaned up or kept: it was most likely incidental. Later, tap Cursor or ABC mid-dictation to pause: what you've said is finished so you can work on it with the cursor pad or the keyboard (the Voice pill says ), and listening starts again at the cursor when you close the pad or keyboard (leaving the field or switching to Gboard doesn't). Press the app's own Send mid-dictation and the box empties: Voicebar stops listening and drops what you'd said (it went with the message) rather than typing it back in. Keys keep working while you talk; ↵ is a line break then, never Send. is one of the own keys' choices, and (a tick when it takes) does the same: say what's wrong in the box ("broad should be lowercase", "drop the okay at the end", "it's Siona, not Shauna"). That's instructions, never typed: the pill turns pink and shows , and when you tap it (or go quiet) the cleanup's model gets everything in the box as it is now (dictated, typed or edited by hand, up to 8,000 characters) with your instructions, and the fixed text replaces the box's text (or goes on the clipboard if the box changed meanwhile). Say nothing and it just takes a closer look. Password fields are never sent. to finish (always polished); to finish and then send, to throw it away (the pill squeezes aside over Send or Cancel; past the mark it says ). Sending uses the field's own Send/Go/Search action when it offers one, otherwise the app's Send button, pressed through the accessibility service (or a note if there's none). Before a dictation is sent, (on by default, Settings ▸ Cleanup ▸ Sending) has the cleanup's model give the whole message one last read for what pauses leave behind: a capital mid-sentence, a sentence split in two, a stray word like a trailing "okay"; its fix is used only if it dropped at most four words and added or reworded none. (Pausing to think is fine anyway: the words after a pause go to the cleanup without the recognizer's capital, and a cleanup at a pause is told you'll go on, so it doesn't end your thought with a period.) With checked (Settings ▸ Cleanup ▸ Sending), an app whose Send button has no label still works: as you swipe, alongside the cleanup, the cleanup's own model (Claude Haiku or GPT-6 Luna) is shown a picture of the buttons to the right of the message box, numbered, and picks the one that sends; it's pressed only if the usual search finds nothing, and if the message leaves the box, that button is remembered for the app (by its view id, or its place from the right). A remembered button that stops working is forgotten, so an app update that moves things gets a fresh look. A hint shows as dictation starts and after 3 s of quiet: round ✕ and send handles in the pill's ends with ‹‹ ›› pointing at them, and between. Polishing shows in the pill, the keys ghosted until the text is in. lives in ☰, with a preview of what's on the clipboard.

Typing with Gboard

ABC pops up Voicebar's keyboard: a full QWERTY that takes over the bar's row too, so its bottom row (?123, ⋮, comma, space, period, ↵) is where your thumbs already are. Pull the handle at the top down to put the voice bar back (or flick it). ⋮ opens the bar's menu over the keyboard, which stays up; hold ⋮ for Gboard. Suggestions across the top finish the word you're typing (your own words first, then the 60,000 most common English words). Swipe to type: slide across a word's letters without lifting; the trail follows your finger, the word goes in with a space before it, and the strip offers the other words it might have been. ⌫ right after a swipe deletes the whole word (and the space the swipe put before it), as Gboard does; once you've typed, picked a suggestion or moved the cursor, ⌫ goes back to one letter at a time. It runs FUTO Swipe's open models on the phone with Voicebar's own decoder (Powered by FUTO Swipe), in a few milliseconds and with nothing sent anywhere; a tap puts one in with a space, and punctuation right after takes the space's place. Capitals come at the start of a sentence; tap shift for one, double-tap it for caps lock. Hold a key for its accents (the top row's numbers first) and slide to one; hold ⌫ to keep deleting; slide on space to move the cursor; two spaces type a period; a second thumb types the first thumb's key at once. ?123 and =< have numbers and symbols, and ↵ shows the field's action (send, search, go). On a space it fixes a lone "i" and missing apostrophes ("dont"); ⌫ right after puts back what you typed. Starting to talk puts it away too. In Settings ▸ Bar and Gboard, ABC can swap the two: tap for Gboard, hold for Voicebar's keyboard.

Hold ABC to switch the field to Gboard. The accessibility service Voicebar: way back from Gboard (Android 11+; on 13+ a sideloaded app first needs Allow restricted settings in its App info) draws a way back while Gboard is open, in one of three styles:

• Bar on top of Gboard (default): a strip in Gboard's own background color (sampled from the screen), joined to its top edge, with a tonal Tap to return pill.
• In Gboard's voice-key spot: covers Gboard's mic and takes the tap; with Gboard's voice key off it sits at the right end of Gboard's top row, sliding left past any round Gboard button (the AI writing key) rather than covering it.
• Floating mic bubble: draggable; remembers its spot relative to the keyboard's top.

Any of them switches back and starts dictation right away. If the service is off, ABC says so first (tap the message to open Accessibility settings, or tap ABC again to switch anyway). The service reads only Gboard's window: its position, its key layout and its background color.

Voice typing

Listen as soon as a text box opens (off by default): dictation starts by itself whenever a text field opens Voicebar (not in password fields).

With a physical keyboard (Bluetooth, USB or a keyboard case): the bar still opens, even though Android normally hides on-screen keyboards then. It starts listening by itself only when the phone is in your hand (half a second of the gyroscope tells a hand's tremor from a phone lying still at a desk); lying still, it says so once ("a keyboard is connected and the phone is lying still") and waits for a tap. Tap Voice, say “voice bar, listen”, or use Gboard's way-back button to dictate.

While you're dictating (or saying what to fix), the screen stays on, so a long dictation isn't cut off by the screen dimming or locking.

When the cleanup fails (a reply that isn't the cleaned text or is far longer than what was said, the service overloaded or rate-limited, a timeout, a dropped connection), Voicebar asks once more before giving up; only a second failure shows the warning. A wrong key or a bad request isn't asked again, and the cleanup at a pause mid-dictation isn't either (the one at the end comes anyway). Fixing the box by voice asks again after a reply that isn't usable. If you swiped to send and the cleanup still fails, the message isn't sent: it's typed as spoken, for a look, and the warning says it wasn't sent.

If Voice is tapped while the app hasn't connected a text box to the keyboard (it happens: the keyboard is up, but there's nothing to type into), Voicebar first asks the app to focus its text box again (the one it was last in, else the focused one, else the lowest on screen) through the accessibility service, waits up to a second for it, then listens into it. A dictation that finishes with no text box gets the same try, for up to 1.5 s, before its words go to the clipboard. Settings ▸ Logs shows each try ("field: …").

Type as you speak (default): the words go straight into the text field as the keyboard's draft and are rewritten in place as the recognizer firms them up; the polished text replaces the draft in one step.

• Press a key mid-dictation and what's typed so far stays; later words continue after the key, and only they get polished.
• Tap in the text or select something mid-dictation and the dictation stops right there and finishes in place: the polish replaces the draft while your cursor or selection stays exactly where you put it. The dictation never pulls the cursor back.
• If the app ends the draft itself, the dictation stops too; the polish lands only if the raw words are still right before the cursor, otherwise your text is left alone.
• Hiding the keyboard or changing fields mid-dictation keeps what was typed.
• Password fields and raw-key fields (terminals) fall back to the preview mode.

Preview mode (setting off): words stream into the voice pill and are typed when you finish.

• Speech to text: ElevenLabs Scribe v2 Realtime over a WebSocket (16 kHz PCM; manual commits at a pause and at the end).
• Cleanup (optional), in two stages: hands-free dictation that goes quiet for half the silence timeout (at most 5 s; 5 s of the default 10) commits what's been said with ElevenLabs and sends it to the Claude Messages API (default claude-haiku-5-5, effort low) while the mic stays open; the cleaned text replaces the raw words in the draft. Keep talking and the next cleanup (another pause, ✓, or the timeout) sends the cleaned text and the new raw words marked apart (<cleaned> / <raw>), so spoken corrections across the pause ("Tuesday… no, Wednesday") still resolve. If nothing was said after the last pause, finishing types the cleaned text with no further request. Any failure or timeout falls back to what's there: cleaned part plus raw words.
• Keys: your own ElevenLabs and Claude API keys, entered in settings, encrypted with an Android Keystore AES-GCM key and excluded from backups.
• Swappable: voice/Transcriber and voice/Cleaner are interfaces; another provider is one class each.
• Grant the microphone from the settings screen (a keyboard can't show a permission prompt itself).

Settings (open the app) use Material 3 with your wallpaper colors: setup steps (turn on Voicebar, switch to it, turn on Gboard, turn on the way back) until everything is ready, then Bar and Gboard, Voice typing, Text and language, a short guide and a text box to try it in.

./gradlew testDebugUnitTest -Prender renders the bar's states, the color schemes and the settings pages into app/build/previews/, for checking the look without a phone.

Icons: gear and cancel from Material Symbols (Apache 2.0); the keyboard-switch glyph is AOSP's ic_ime_switcher (Apache 2.0).

Estimated cost

Settings → Voice typing → Estimated cost shows what dictation cost over the past 24 hours, 7 days and 30 days, split into Scribe and Claude. Voicebar counts what it actually sent: the seconds of audio streamed to Scribe v2 Realtime (priced at $0.39 per hour), and the input and output tokens Claude reports back in each cleanup reply (priced per model at list rates; Haiku 5.5 is $0.10 / $0.50 per million). The counts are kept on the phone in hourly buckets for 31 days; no text or audio is stored. It's an estimate at list prices: plan allowances, free hours and credits aren't taken into account, so the providers' dashboards are the real bill. A model with no known price has its tokens counted but not priced.

Language processing

• Millennial mode (off by default): a dictated message never ends with a period. Question marks, exclamation marks, an ellipsis and periods between sentences stay. Dictate more right after such a message and its period comes back before the new words (and goes away again if that follow-up is thrown out or hears nothing).

Hands-free mode

Say “voice bar, hands free” (while the bar or the Gboard button is up, or anywhere with Listen for it everywhere on), or use the Hands-free Quick Settings tile, and control the phone by voice. It runs in the accessibility service, which reads every app's screen and acts on it with real taps and gestures (canPerformGestures).

• Listening: ElevenLabs streams while it's on and ends each command at a pause (commit_strategy=vad). The pill in the status bar (left of the camera) shows what it heard and what it's doing; tap it, or say “stop”, to end. It turns off after 5 minutes without a command, or when the screen goes off.
• On the phone: tap / click / long press name or number · show / hide numbers (badges on everything tappable, like Voice Access) · scroll up/down/left/right, with how far (“a bit”, “half a screen”, “one full screen”, “two pages”, “a lot”, “to the top”, “all the way down”, “back to top”; exact amounts are dragged with a finger that holds still before lifting, so nothing flings past) · page up/down · swipe left/right · zoom in/out · type words · clear · focus the search box · open app or thing on screen · back · home · recent apps · notifications · quick settings · lock · screenshot · cursor up/down/left/right (a bit) · cursor to name · tap.
• Anything else goes to Claude (Haiku by default) with the screen's elements (role, label, position, what each can do) and a fixed set of actions: tap, type, scroll, swipe, drag, a system action, open an app, or done. Apps that hide their UI from accessibility (maps, games) get a screenshot too. “Go ahead and …” runs a goal of up to 5 steps, planned one at a time on the screen each step leaves.
• The cursor glides to each target before acting and ripples where it taps; scrolls and swipes draw a trail.
• Asks first: anything that sends, posts, pays, buys, deletes or submits (Claude marks it, and the phone checks button labels) waits for “yes”, with the question beside the cursor. “No”, silence (15 s) or another command cancels.
• Typing uses the accessibility set-text action (adds to what's in the field, with your dictionary's fixes). Password fields are never read.

Cost while on: ElevenLabs streaming (about $0.39 an hour) plus a fraction of a cent per Claude-planned command. Settings → Voice typing → Hands-free.

Parakeet on the phone

Settings › Voice typing › Speech recognition chooses where speech becomes text: ElevenLabs Scribe (cloud, live words, your key) or Parakeet on this phone: NVIDIA's Parakeet Unified 0.6B (April 2026, English, with punctuation and capitals), run by sherpa-onnx with no network and no key. The model (663 MB) downloads once, resumably, from a pinned Hugging Face revision of the sherpa-onnx export, and is checked against known sizes and SHA-256 sums; Remove frees the space. The setup wizard offers it as an alternative to an ElevenLabs key.

Parakeet Unified's true streaming exports re-run a context window per chunk and measured slower than real time even on an M1 Max, so Voicebar uses the regular model with on-device voice detection (Silero VAD, bundled): each stretch of speech is read once at the pause and committed, and between pauses the stretch so far is re-read about every 0.8 s (less often on slower phones) so words still show while you talk. On an M1 Max that whole loop costs about a tenth of real time. Hands-free mode uses it the same way, one command per utterance. The model loads when the keyboard opens (if chosen) and is let go after three idle minutes. Cleanup still uses Claude when you add a key; keyterms and screen context for recognition are ElevenLabs-only.

Re-dictating inside text

When dictation starts with text already in the field, the cleanup gets the text either side of the cursor (up to 1,000 characters before, 400 after; never from password fields) as <before> and <after>, and is told to return only the new words, fitted to that spot: no capital mid-sentence, no closing period when the sentence carries on after the cursor, the surrounding style kept, and the whole sentence used to pick the right sound-alike. A reply that echoes the field anyway has that part taken back off. The words just before the cursor also go to ElevenLabs as the lead-in for recognition (read from the field itself, so this works with screen context off), and when the new words land against a word after the cursor, a space goes between them.

Ignore children's voices

Settings › Voice typing › Other voices › Ignore children's voices silences a small child talking between your words, with nothing to set up and no model to download. A young child's voice sits more than an octave above an adult man's (roughly 250–400 Hz against 85–155 Hz), so pitch alone tells them apart. Audio goes through a short-time Fourier transform on the phone (64 ms frames every 16 ms). Each frame's pitch is the lowest one whose harmonics stand out over the troughs between them, checked against being an adult's bright 2nd–6th harmonic, judged over the last few frames, and required to glide the way a voice does. A stretch that's only a child (no adult heard in the last ~130 ms) is replaced with silence, held ~220 ms past the child's last voiced sound for the consonants. About 50 ms of delay; it runs before Only my voice when both are on.

How high counts as a child sets the line from 260 Hz (Gentle) down to 190 Hz (Strict); stronger catches older children but may cut adults with high voices. When a child talks over you at the same moment, the audio can't be pulled apart this way, so cleanup is also told to drop words that are plainly a child cutting in.

Tested with synthetic voices (macOS's Daniel as the adult; Samantha raised to a child's pitch and vocal-tract size) through Parakeet: a child talking alone or between the adult's sentences went from 28–29 transcribed words to none, the adult's sentences were transcribed the same as unfiltered, and overlapping speech was unchanged.

Only my voice

Settings › Voice typing › Other voices › Only my voice keeps other people out of your dictation and hands-free commands: someone talking between your sentences, a TV, a conversation nearby. It needs a speaker model, 3D-Speaker ERes2Net (26 MB, downloaded when you turn it on, pinned and SHA-256-checked like Parakeet), and works one of two ways:

• With a voiceprint (Your voiceprint, made on the phone from about 20 seconds of you reading four sentences aloud), it always knows your voice, in dictation and hands-free.
• Without one, in a dictation you start yourself (a tap or hold on Voice), the first voice at the phone is the one kept: whoever started it is the one holding the phone. The first 1.5 seconds of speech at least as loud as you usually dictate (your learned level, less about 6 dB) become that dictation's reference, and from then on it's judged like a voiceprint, with a little leeway (0.04) for the shorter sample. Before that, quieter speech (a TV, a conversation across the room) is silenced, so it can't be taken for you. Each dictation starts afresh. Hands-free and auto-started dictations aren't filtered this way, since nobody tapped to start them.

While you dictate, audio goes through in half-second blocks. Each block with speech in it (Silero VAD) is compared with your voiceprint using the last second of speech; a block that doesn't sound like you is replaced with silence before ElevenLabs or Parakeet hears it. That adds about half a second, plus up to another half second at the start of speech while there's enough to judge. A pause of half a second starts the judging afresh, so a new speaker isn't judged by your last words, and a word too short to judge ("yes") is let through. Strength (Gentle to Strict) sets how closely speech must match: stronger silences more of other voices but may drop some of yours in noise or when you whisper. Check it, on the voiceprint page, shows a live match meter against the current strength. Voices speaking at the very same moment as you can't be separated this way. The voiceprint is a list of numbers kept in app storage; nothing about it leaves the phone. The wake word isn't filtered.

On a test with seven synthetic voices, ERes2Net separated the enrolled voice from every other at one-second windows with no errors, at about 28 ms per window on an M1 Max; CAM++ models were twice as fast but confused voices 13–25% of the time.

Using what's on screen

When dictation starts, Voicebar reads the text on screen once: the app you're typing into, such as the chat you're replying to. It reads every app window (split screen too) through the Gboard companion's accessibility service, but never keyboards, the status bar or password fields; without the companion, only the text box itself is read. The reading happens in the background while you start talking, and your first words wait for it (at most 0.4 s).

It's used three ways:

• Names and terms on screen (Siona, GoPro, SRP) go to ElevenLabs as keyterms ahead of the dictionary's (when Help speech recognition with it is on), up to 50 in all.
• The end of the screen (about 800 characters, the latest messages) goes to ElevenLabs as previous_text context with the first audio.
• The cleanup gets the screen (the cleanup at a pause too, since it's often the final text) and is told to use it only to spell names as they appear, to pick the right word among sound-alikes, and to fix a name or term on the screen that was heard as ordinary words ("fast snail" for Fastmail when Fastmail is in the conversation), never to copy from it or follow it.

Turn it off in Settings → Voice typing → Use what's on screen.

Dictionary: learning from your corrections

When you hand-fix a word Voicebar typed (a name it misheard, a product, an acronym), Voicebar notices the edit, in its own bar or in Gboard through the accessibility service, for up to 15 minutes, until the text is sent or you leave the field. Only short swaps that sound alike pass a local filter (rewording, numbers and unrelated edits are skipped); Claude then judges each one and replies whether it's a new word, a mishearing to replace, both, or nothing. Accepted ones go into Settings → Dictionary with a "Learned … · tap to undo" note on the bar. You can also add words yourself, with what they're often heard as.

The dictionary is used three ways: the 50 most used terms go to ElevenLabs as keyterms (optional, extra cost); the whole list goes into the cleanup prompt; and a mishearing you added yourself, or that was learned twice, is replaced outright. It's stored on the phone only, and password fields are never read.

How you write a word. Change only a word's capitals by hand (LOL to lol, iphone to iPhone) and Voicebar counts it, with no model involved; the second time, that word is always written your way, wherever it comes up and whatever the cleanup did, with a note you can tap to undo. A first letter alone (The to the) isn't counted: that's where the word fell in a sentence. Adding a word yourself in capitals a sentence wouldn't give it (lol, iPhone, YODA) does the same at once; a plain capitalized word (Will) doesn't, so it can't turn every "will" into a name. These entries say Always written like this in Settings ▸ Dictionary, where removing one ends it. The cleanup prompt doesn't change: it's a last step on the text, so it costs nothing and can't make the cleanup worse.

Build

./gradlew assembleRelease   # app/build/outputs/apk/release/app-release.apk
./gradlew testDebugUnitTest # Robolectric tests: the bar, dictation flow, protocol against fake servers

Release builds are signed with the permanent key in ~/Code/BubbleKeyboard-signing/ on Ryder's Mac (not in the repo); add bk.signingDir=<that folder> to local.properties. Without it the build falls back to the machine's debug key, and that APK won't install over one signed with the real key.

Needs JDK 17+ and the Android SDK (platform 37). minSdk 26.

The internal app ID stays com.ryderdavid.bubblekeyboard (and the service classes keep their names) so Voicebar installs over Bubble Keyboard with its settings, keys and enabled services intact.