Suflyor

---

Suflyor (Russian суфлёр, "prompter") is a free, open-source teleprompter app for Android that follows your voice. It floats over Instagram, TikTok or the camera app and scrolls the script as you speak, so you never chase the text: a voice-activated teleprompter for Reels, Shorts and TikToks. It works fully offline, and the app has no internet permission at all.

It is made for creators who hold the phone, read a couple of lines to the lens, look back at the text and carry on. If you stumble and start a sentence again, the text goes back with you. If you skip a paragraph, it catches up.

[!NOTE] Follows your voice in 13 languages. English, Russian and Turkish are built in; Portuguese, Spanish, German, French, Italian, Dutch, Swedish, Hebrew, Korean and Chinese are speech packs you download once. Suflyor picks the language from your script's text, or you set it in Settings. The interface is in English, Russian, Turkish and Brazilian Portuguese and follows your phone's language.

Screenshots

Library	Script	Editor	Voice following
All your scripts, imported from almost any format	Preview, reading time and one tap to start over the camera	Plain text with **emphasis**, # headings and [notes]	Ten speech languages, picked from the script or set by hand

Features

Follows your voice

• On-device streaming speech recognition for English, Russian and Turkish (sherpa-onnx), plus Portuguese, Spanish, German, French, Italian, Dutch, Swedish, Hebrew, Korean and Chinese as speech packs you download once (27 to 123 MB each, see Speech packs). No cloud, no account, no delay from the network.
• Understands how people read English: contractions (don't, it's), numbers written as digits but said as words, abbreviations (Mr., Dr.), acronyms (AI, CEO) and "um"s.
• Searches the whole script, not only the next few words. Rare words count more than common ones, so it doesn't jump on every "and".
• Re-reading a line brings the text back to it after about two words. A skipped block is found after three or four.
• Pauses and off-script talk leave the text where it is.
• Lines you have read dim away. [Stage directions in brackets] stay visible but are never expected to be spoken.

Floats over any camera app

• Keeps listening while Instagram, TikTok or the stock camera records video. The mic isn't taken away from the prompter, and your reel still gets the sound.
• The text starts right under the front camera, so your eyes stay near the lens. The controls sit at the bottom.
• Lock mode: touches pass through the window to the camera's own buttons.
• Landscape: the window moves next to the lens and turns the text, even when the camera app keeps the screen in portrait.
• Quick Settings tile: starts the prompter over whatever app is open.
• Aa in the window: text size, lines, background and speed, changed right over the camera.

Remotes and scrolling

• Volume keys, Bluetooth selfie remotes, rings, clickers and keyboards. Each key can be assigned in Settings.
• Timed scrolling at a set speed, with a 3-2-1 countdown, when you'd rather not use your voice.
• Scrolling by sound: the text moves at that speed while you talk and waits while you're quiet. It works for a script in any language.
• Full-screen prompter inside the app, with text up to 80 sp, remote keys and a mirror mode for teleprompter glass.

Your scripts are yours

• Share a script's text, duplicate it to try another version, or save the whole library to a zip of Markdown files and restore it on another phone.
• The editor counts words and the length of the take as you type.

Imports almost anything

• TXT (UTF-8, UTF-16, Windows-1251, KOI8-R), Markdown and Obsidian notes, DOCX, ODT, RTF, HTML and PDF (PDF needs Android 15, or Android 12–14 with recent Google Play system updates).
• Open from other apps with Share or Open with, paste from the clipboard, or type in the editor.
• Long paragraphs are split into short phrases you can read in one breath.

How it works

flowchart LR
    Mic["Microphone"] --> ASR["sherpa-onnx<br/>streaming recognizer<br/>(on the phone)"]
    ASR -->|"last recognized words"| Tracker["Script tracker<br/>fuzzy local alignment<br/>over the whole script"]
    Tracker -->|"position"| View["Prompter window<br/>over the camera app"]
    Script["Your script"] --> Tracker

The tracker aligns the last few recognized words against the entire script. It keeps only matches that end on one of the last two words you said (the newest counts more), and it weighs every word by how rare it is in your script. Small steps forward need a single good match. Going back to a nearby line needs about two words from its start. A far jump needs at least three words and a clear lead over every other place in the script; unless four or more rare words match, it also waits for the next update to agree.

Privacy

Everything stays on your phone. The app does not have the INTERNET permission, so Android won't let it connect anywhere, and you can check this yourself.

• Audio is used only while a prompter session is running. It is never recorded, stored or sent.
• The accessibility service is active only during a session. It cannot read screen content. It keeps the microphone working while other apps record, shows the floating window, and handles volume and remote-control keys (on by default with a preset set of keys; you can turn this off or reassign the keys in Settings).
• No analytics, no ads, no accounts, no trackers.

The details are in PRIVACY.md.

Install

Requirements: Android 10 or newer on a 64-bit ARM phone (almost every phone since 2017), about 135 MB of space: the speech models are inside the app. Tested on a Samsung Galaxy S24 FE with One UI 6.1.1 (Android 14), see tested phones.

1. Download Suflyor-<version>.apk from Releases and open it. Allow your browser or file manager to install apps when asked.
   • Lite version: Suflyor-<version>-lite.apk (about 30 MB) has no speech models inside; you then download the pack of your language once, Russian, English and Turkish too (see Speech packs). It is published as a separate pre-release "Suflyor lite" next to each version, so Obtainium keeps installing the full APK. Both install over each other and keep your scripts.
2. Open Suflyor and allow the microphone.
3. Turn on the accessibility service “Suflyor: floating window and voice” (on a Russian phone: «Суфлёр: окно поверх и голос»). The app takes you to the right screen.
   • On Android 13 and newer, apps installed from a file can't use accessibility at first, and the switch is greyed out. Go to Settings → Apps → Suflyor → ⋮ → Allow restricted settings, then turn the service on again.
4. On a script, choose Instagram, TikTok or Camera under the Over the camera button, tap the button and start reading.

Language: the app follows your phone's language; on Android 13 and newer you can pick English, Russian, Turkish or Brazilian Portuguese just for Suflyor in Settings → About → App language. The speech language is separate: Auto picks it for each script from its letters; pick it for one script with the Speech button on the script screen, or change the default in Settings → Voice following.

Updates: new versions install over the old one. With Obtainium you get them automatically from this repository.

If Android says the developer isn't verified

Google is rolling out developer verification. In 2026 it doesn't affect installs from GitHub or Obtainium. From 2027, phones with Google services install apps from developers who are not yet verified only after a one-time setting (the "advanced flow"):

1. Turn on Developer options: Settings → About phone → tap Build number seven times (on Samsung: About phone → Software information → Build number).
2. In Developer options, turn on Allow apps from unverified developers and confirm with your screen lock. Android also asks you to confirm that nobody is guiding you through this: it is a protection against scams.
3. Restart the phone and wait 24 hours. This safety delay happens once.
4. Open the setting again, confirm, and choose indefinitely (or 7 days).
5. Install the APK and tap Install anyway.

Also works: installing from a computer with adb install Suflyor-<version>.apk, with no waiting. Phones without Google services (LineageOS without GApps, Huawei) are not affected.

If Play Protect says "App blocked to protect your device"

In some countries (India, Brazil, Vietnam, Thailand, Singapore, the Philippines, Kenya, Nigeria, South Africa and others) Google Play Protect has enhanced fraud protection. It blocks any app that uses an accessibility service when you install it from a browser, a messenger or a file manager. The message says "This app can request access to sensitive data", then the installer shows "App not installed". The APK is fine: Suflyor needs the accessibility service so that Android keeps its microphone on while the camera records, and the check looks only at that.

Any of these works:

1. Pause Play Protect for the install: Play Store → your profile picture → Play Protect → the gear icon → turn off Scan apps with Play Protect. Install the APK, then turn scanning back on.
2. Install from a computer with adb install Suflyor-<version>.apk.
3. Install through Obtainium. Google says the check skips app stores, and Obtainium installs as one. It also brings updates. If it still gets blocked, use option 1.

After the install, turn on the accessibility service as in step 3 above.

Speech packs

Russian, English and Turkish models are inside the full app (in the lite version they are packs too: Russian 23 MB, English 60 MB, Turkish 6 MB). Portuguese, Spanish, German, French, Italian, Dutch, Swedish, Hebrew, Korean and Chinese are separate downloads of 27 to 123 MB each, so the app stays small for everyone else.

1. Open Settings → Voice following → Speech languages and tap Download next to the language. Your browser downloads suflyor-speech-<language>-1.zip from this repository's speech-packs releases.
2. Tap Install downloaded pack and pick the file, or open the file from your browser's downloads with Suflyor.

Suflyor itself still has no internet permission: the browser downloads the file, and the app only reads what you give it. A pack is installed only if its SHA-256 matches the list inside the signed APK, so a changed or broken file is refused. You can delete the zip after the install.

Would you rather have one-tap downloads and updates inside the app, at the cost of an internet permission? Vote in the poll.

Most packs are Kroko community models (CC BY-SA); Korean and Chinese are k2-fsa Zipformer models (Apache-2.0). Kroko recognizes speech in steps of 1.28 s, so with those packs the text follows about a second later than with the built-in languages; Korean and Chinese step every 0.32 s. Measured on audiobook clips from Multilingual LibriSpeech (Swedish: FLEURS): about 5% wrong words in Spanish, 8% in Portuguese and German, 10% in French, 11% in Swedish, 13% in Dutch, 19% in Italian and 32% in Hebrew (FLEURS, read Wikipedia sentences full of names). Chinese and Korean recognizers print no spaces, so Suflyor follows them character by character and matches characters that sound alike: 18% wrong characters in Chinese, 26% wrong syllables in Korean. On a Galaxy S24 FE a pack loads in about 2 s, and the tracker followed every test text to the end.

Verify your download

Release APKs are built by GitHub Actions from the tagged source. Each release carries a SHA256SUMS.txt file and a signed build-provenance attestation. All releases are signed with the same key; the SHA-256 fingerprint of its certificate is:

77:F3:94:32:F5:05:FB:B7:8B:85:E7:37:CD:00:21:0A:44:C0:7A:13:2C:AC:FF:81:FE:9D:9D:79:CD:E9:A9:06

# The certificate must match the SHA-256 above:
apksigner verify --print-certs Suflyor-*.apk
# Proves the APK was built by this repository's workflow:
gh attestation verify Suflyor-*.apk --owner olerast67

AppVerifier can check the certificate on the phone.

Build from source

You need JDK 17 and the Android SDK (platform 36).

git clone https://github.com/olerast67/voice-teleprompter-android.git
cd suflyor
./gradlew assembleDebug testDebugUnitTest

• Downloads on first build. The speech library (sherpa-onnx 1.13.8, ~50 MB), the Russian model (~28 MB), the English model (~73 MB) and the Turkish model (~7 MB) are not stored in git. The fetchSpeechAssets task downloads them from pinned revisions and checks every file against a pinned SHA-256.
• Release builds (./gradlew assembleRelease) are signed only when a key is configured, through keystore.properties or the SUFLYOR_* environment variables. Without a key they come out unsigned. See RELEASING.md.
• On Windows, .\build.ps1 and .\install.ps1 read sdk.dir and jdk.dir from local.properties.

Path	What's there
app/src/main/java/.../ui	Jetpack Compose screens: library, script, rehearsal, editor, settings, journal
.../overlay	Floating window, accessibility service, prompter view, Quick Settings tile
.../track/ScriptTracker.kt	Voice-following: word alignment against the script
.../speech	Microphone capture and sherpa-onnx recognizer
.../session	Session engine and foreground service
.../doc	Importers: TXT, Markdown, DOCX, ODT, RTF, HTML, PDF
.../script	Word normalization and phrase layout
app/src/test	Unit tests for the tracker (with a random stress test) and importers

Tested phones

Whether the voice keeps being heard while a camera app records depends on the phone's firmware. What has been checked so far:

Phone	Android	Instagram	Samsung Camera	TikTok
Samsung Galaxy S24 FE, One UI 6.1.1	14	works	works	not checked

Tried it on another phone? Open an issue with the model, the Android version, which apps you recorded in and whether the text followed you. "Everything works" is a useful report too.

Roadmap

• More speech packs: Bengali has a streaming model and needs its own word rules. Japanese can be followed by characters like Chinese but has no small streaming model; Polish and Vietnamese wait for one too.

• One large pack for many languages: NVIDIA's Nemotron streaming model covers 35 languages (Japanese, Hindi, Ukrainian, Polish, Vietnamese and others). It keeps up with speech on a Galaxy S24 FE, but it is 680 MB and needs six times the computing of the current packs, so it waits for a test on a mid-range phone. The measurements are in docs/speech-models.md.

Ideas are welcome in Issues. Should the app get internet access for one-tap downloads and updates? Vote in the poll.

Found a bug? Open an issue or use the form on olerast.com: Report a bug. The form is in Russian and needs no GitHub account.

FAQ

Yes: Suflyor listens through the microphone and scrolls to the words you are saying. It is free and open source, and the speech recognition runs on the phone.

Yes. Suflyor floats over any app, and its text starts right under the front camera. Android normally takes the microphone away from other apps while a camera app records; Suflyor's accessibility service is what lets it keep listening (the reel still gets its sound).

It works fully offline. The app has no internet permission, so it can't send anything anywhere; audio is never recorded or stored. See PRIVACY.md.

English, Russian and Turkish speech are built in. Portuguese, Spanish, German, French, Italian, Dutch, Swedish, Hebrew, Korean and Chinese come as speech packs you download once in Settings → Voice following. The language is picked automatically from the script or set in Settings. The interface is in English, Russian, Turkish and Brazilian Portuguese.

Yes: volume keys, Bluetooth selfie remotes, rings, presentation clickers and keyboards can pause, go back or forward a line. Keys can be reassigned. Timed scrolling with a 3-2-1 countdown is there too.

Free, with no ads, no accounts and no paid features. The code is GPL-3.0. If it helps you, you can support it.

Android 10 or newer on a 64-bit ARM phone (almost every phone since 2017). There is no iOS version: iOS doesn't let an app float a window over other apps the way Suflyor needs.

Support the project

Suflyor is free and always will be: no ads, no paywalls, no "pro" version. If it saves you retakes, you can support its development. Every donation goes into more languages and more devices to test on.

 

Starring the repository and telling other creators about it helps too.

Contributing

Bug reports from different phones are the most valuable help. Attach the in-app log: Settings → About → Log → Share. See CONTRIBUTING.md for building, code style and adding a language. Report security problems privately, as described in SECURITY.md.

Credits

• sherpa-onnx by the k2-fsa team: streaming speech recognition (Apache-2.0).
• vosk-model-small-streaming-ru by Alpha Cephei: the Russian model (Apache-2.0).
• The icefall streaming Zipformer trained on LibriSpeech: the English model (Apache-2.0).
• seda-v0.2-tiny by atasoglu: the Turkish model (CC BY-SA 4.0).
• Kroko ASR community models by Banafo: the speech packs (CC BY-SA). Turkish support was contributed by @Higgstor.
• ONNX Runtime (MIT), Jetpack Compose (Apache-2.0).
• The full list is in THIRD_PARTY_NOTICES.md.

How this was made: the author designed the app, directed it and tested it on a real phone. Much of the code was written with an AI coding assistant (Claude by Anthropic). The voice tracker (in both languages) and the importers are covered by unit tests in app/src/test.

License

GNU General Public License v3.0. You may use, study, share and modify Suflyor. Modified versions you distribute must stay open source under the same license.