SandFly is a screen reader for Android. It is a fork of Backtalk, which is a fork of Google's TalkBack. It is not affiliated with Google or with Backtalk.
SandFly runs on phones and tablets with Android 8 or later, on Wear OS watches, and on Android TV.
Download sandfly.apk from the latest release and open it on the phone. The first time, Android asks you to allow your browser or file manager to install apps. For a Wear OS watch, install sandfly-wear.apk with adb.
SandFly installs as io.github.aaron_gh.sandfly, next to Google's TalkBack and Backtalk. Each app has its own settings.
Go to Settings > Accessibility. SandFly is listed as SandFly and is off. Turn off any other screen reader first, then turn on SandFly.
SandFly checks for a new release when it starts and about once a day. When there is one, it shows a notification with the list of changes. Tap the notification to download and install it. The first time, Android asks you to allow SandFly to install apps.
To check now, go to Check for updates in SandFly settings, which also shows the version and build number. To stop automatic checks, turn off Automatically check for updates.
There are stable releases and development builds:
- Stable releases are on the releases page, named by version, such as SandFly 0.1.0.
- Development builds are on the dev release page, which each change replaces. Turn on Development builds in SandFly settings to get them. SandFly then checks about once an hour, and offers whichever is newer, so a development build is offered the next stable release.
sandfly.apk and sandfly-wear.apk each update to builds of their own kind. On a watch, the update downloads over Wi-Fi when it is available, with its progress in a notification.
The last build of Backtalk Aaron offers SandFly as an update. To move:
- Install SandFly from the notification, or from Check for updates in Backtalk Aaron. It installs next to Backtalk Aaron, and the accessibility settings open.
- Turn on SandFly. The settings and custom labels of Backtalk Aaron are copied to it. SandFly gets stable releases; turn on Development builds to get development builds too.
- Backtalk Aaron turns itself off, and a notification offers to uninstall it.
On-device AI models are not copied, so download them again in SandFly. Turn on the braille keyboard again for SandFly if you used it.
If you use an accessibility shortcut, grant SandFly permission to change secure settings before you turn it on, and it moves the shortcut to itself: adb shell pm grant io.github.aaron_gh.sandfly android.permission.WRITE_SECURE_SETTINGS. Otherwise, set the shortcut again for SandFly.
SandFly speaks sooner after each swipe and screen change, and can play its sounds and speech through a low-latency audio path. It adds a two-finger rotor, edge gestures and a two-finger scrub as in VoiceOver, a circle menu, Pause SandFly, direct touch for audio games, sound themes with control sounds in 3D, a vibration for every sound, a status gesture, voice profiles, settings for how notifications are read, its own notification list and quick settings panel, image and screen descriptions from several AI services or a model on the phone, and braille keyboard dots that follow how you hold the device. It also changes some defaults and fixes problems in TalkBack. For the full list, see Differences from TalkBack.
Scripts written in Lua can change how SandFly works in one app or in every app. See Writing SandFly scripts. Other apps can speak in the user's SandFly voice, through the speech API, and give SandFly a braille display it can't reach on its own, such as a notetaker's, through the braille display API.
Two settings need permissions that Android only grants through adb:
- Touch and activation > Fix stuck touch exploration, which is on by default, needs the DUMP permission to fix touch exploration when Android leaves a touch down.
- Sound and vibration > Speaker when away from your ear needs permission to manage calls.
With Shizuku installed and running, SandFly grants these itself: DUMP when SandFly starts, and the call permission when you turn the setting on. The first time, Shizuku asks whether to allow SandFly. Without Shizuku, each setting shows the adb command, and its Copy command action copies it.
Describe image and Describe screen can use an AI service with your own API key, or a model that runs on the phone. Google's source release does not include working Gemini support, so without one of these, Describe image only reads text in images, and Describe screen does not work. SandFly adds its own support for Describe screen, including follow-up questions.
- Open SandFly settings, then Automatic descriptions, and turn on Detailed image descriptions.
- Under AI services, open the settings of a service: Gemini, Claude, OpenAI (ChatGPT), Mistral, Groq, OpenRouter, Ollama Cloud, or Other service for another service with an OpenAI-compatible API, whose address you enter.
- Choose Get a key, where the service has one, to open its website and create a key. Paste the key into the API key field, and choose Save.
- Choose a model from the list that the service gives for your key, or enter its name.
- If no other service is in use, SandFly uses this one as soon as it is set up. AI service in Automatic descriptions shows the service in use and lists the services that are set up, to switch between them. None turns AI descriptions off.
A service can't be used until it has an address, a key and a model. When you clear the key or model of the service in use, SandFly uses the next service that is set up, or None. Questions about an image or screen, spoken or typed, go to the service in use.
Images and screenshots that you describe are sent to the service. On Gemini's free tier, Google can use this data to improve its products. With Gemini, when the chosen model is busy or takes more than 15 seconds, SandFly asks Gemini Flash-Lite instead.
If you do not want to use an API key, or you have hit the free tier limit, SandFly can describe images and screens with a Gemma 4 model that runs on your phone. Nothing is sent to Google or anyone else when you use it.
- Open SandFly settings, then Automatic descriptions, then On-device AI.
- Choose a model. Gemma 4 E2B is the one to start with: 2.6 GB, and it needs a phone with about 6 GB of memory. Gemma 4 E4B is 3.7 GB, gives better answers, and needs about 8 GB. The list also has other small vision models from the LiteRT community, from 0.4 GB up, so that you can try them. Those are marked experimental: they are community conversions that the SandFly developers have not tried, and some may not work or may follow the screen description format badly. The list only shows models that your phone has enough memory for, plus any you already have. If your phone does not have enough memory for Gemma 4 E2B, SandFly picks the largest model that fits.
- Choose Download model. It downloads once from Hugging Face and carries on where it stopped if the connection drops. A notification shows the progress. SandFly checks the file against a known SHA-256 hash and deletes it if it does not match. Or choose Use a model file from storage to use a
.litertlmfile that you downloaded yourself, such asgemma-4-E2B-it.litertlmfromlitert-community/gemma-4-E2B-it-litert-lm. SandFly works out which model it is. - Turn on Use on-device AI.
Detailed image descriptions in Automatic descriptions must also be on. While Use on-device AI is on, Describe image and Describe screen use the model on your phone instead of the AI service. Turn it off to go back to the AI service.
Answers take several seconds and use battery, more than the cloud on a mid-range phone. The model loads on the first request and unloads after two idle minutes to free memory. The model runs in its own process, so if the phone runs out of memory, Android stops the model and not SandFly, and SandFly says so. Before it loads a model, SandFly checks that the phone has at least as much free memory as the model's size, and if not, it says there is not enough free memory and does not load it. To test with a model that is too big for your phone, turn on Ignore on-device AI memory limits in Advanced settings > Developer settings. It lists every model, lets you choose any downloaded one, and skips the free memory check. If answers fail or the phone slows down, try the other model, or change Processor. On-device AI needs a 64-bit ARM phone and adds about 22 MB to the app.
Processor chooses where the model runs, and appears only on phones that have more than the CPU: GPU on phones whose vendor lists an OpenCL driver for apps (libOpenCL.so in Android's public library lists), and NPU on chips that have a vision model built for their NPU: Google Tensor G5 and G6 (Pixel 10 and 11) with Gemma 4 E2B (3.1 or 3.3 GB), and Snapdragon 8 Elite (SM8750, as in the Galaxy S25 and Z Fold7) and 8 Elite Gen 5 (SM8850) with FastVLM 0.5B (0.9 GB, Apple research licence). With NPU, the model is that NPU build instead of the chosen one, and the NPU's libraries are downloaded with it, checked against their SHA-256: for Tensor, LiteRT's libLiteRtDispatch_GoogleTensor.so (com.google.ai.edge.litert:litert-npu-runtime-google-tensor 2.3.0, Google's Maven repository); for Snapdragon, LiteRT's libLiteRtDispatch_Qualcomm.so (litert-npu-runtime-qualcomm 2.3.0) and Qualcomm's libQnnHtp.so, libQnnSystem.so and the Hexagon v79 or v81 stub and skel from com.qualcomm.qti:qnn-runtime 2.50.0, the QAIRT version LiteRT 2.3.0 is built with (Maven Central, Qualcomm AI Hub Model License, which allows them in apps). None of them are in the APK or loaded on other phones. MediaTek NPUs have no vision model built for them yet. The speech models built for NPUs (Whisper, Parakeet, Moonshine) need LiteRT's C++ speech runtime, which LiteRT-LM's Android library doesn't include, so dictation doesn't use NPUs.
Live recognition in SandFly settings opens a screen that describes what the rear camera sees. Its modes turn on and off separately: Scenes, Text, Point and Speak (the text above or below a finger in the camera view), People detection and Door detection. With the Live recognition reading control (Android 11 and later), swipe up or down to choose a mode and double-tap to turn it on or off; the screen opens if it isn't open. Double-tap works on the mode the reading control says as soon as it is chosen, until focus moves. On the screen, a two-finger tap stops speech and speaks the next result, and a two-finger scrub closes the screen.
Text is recognized on the phone by ML Kit's Latin-script text recognizer, which Google Play services may download first. The other modes are described by what Describe image uses:
- With Use on-device AI on, the model on the phone describes each frame as soon as it has finished the last one. Nothing leaves the phone. Small models such as FastVLM get short plain questions instead of the full instructions.
- With another AI service in use, such as Claude or OpenAI, a camera image goes to it at most every 4 seconds, through its chat API, once the previous answer is in. SandFly asks once before the first image goes to each service.
- With Gemini in use, a camera image goes at most every 4 seconds to Gemini 3.8 Live (
gemini-3.8-live, Live API over a WebSocket), as soon as the previous answer is in, with the Gemini API key from Gemini settings. SandFly asks once before the first image is sent. Gemini answers with audio, which SandFly drops; the text spoken is Gemini's transcript of it.
Answers are spoken with SandFly's voice, even while other audio plays. The microphone isn't used. An answer from Gemini that comes more than 10 seconds after its image isn't spoken, nor one from the on-device model or another service after 30 seconds. A rejected API key or a used-up quota turns off the modes that use AI. With the on-device model or another service, so do three failed requests in a row; with Gemini, any failure does.
Leaving the screen, locking the phone, pausing SandFly or turning off every mode turns off the camera and closes the Gemini session. People and doors are described from the image only, without distances. Don't rely on the descriptions to move around safely.
SandFly's own strings are translated into German, Afrikaans, Amharic, Arabic, Assamese, Azerbaijani, Belarusian, Bulgarian, Bangla, Bosnian, Catalan, Czech, Danish, Greek, English (United Kingdom, Australia and India), Spanish (Spain and Latin America), Estonian, Basque, Persian, Finnish, Filipino, French (France and Canada), Galician, Hindi, Vietnamese and Simplified Chinese. Strings that are not translated yet, such as most of the notification settings, show in English. Strings from TalkBack keep Google's translations.
SandFly includes the Android TV support from Google's TalkBack, and the same sandfly.apk installs on Android TV and Google TV. It has been tried on an Android 10 TV emulator, but not yet on a real TV.
- Turn on developer options and network debugging on the TV, then connect to it with
adb connectand the TV's IP address. - Install the APK with
adb install sandfly.apk. - Turn on SandFly in the TV's accessibility settings. On most Android TVs, this is Settings > Device Preferences > Accessibility.
To open SandFly settings, choose SandFly settings in the apps on the TV home screen. TVs often don't show notifications, so to update, go to Check for updates in SandFly settings. Pause SandFly isn't available on TVs, because a TV remote often has no way to resume it.
You need JDK 17 or newer, the Android SDK with platform 37, and NDK 27.3.13750724. The Gradle wrapper downloads the correct Gradle version.
Set ANDROID_SDK to your SDK path, then run ./build.sh. This builds the phone and watch APKs, and keeps anything else that's in local.properties, such as a Gemini API key.
Set ANDROID_HOME to your SDK path, then run:
.\gradlew.bat assemblePhoneDebug
.\deploy.ps1 builds the APK and installs it with adb. This script needs PowerShell 7. To install the last build without building again, use -SkipBuild.
Android only installs an update that is signed with the same key as the installed app. Builds from GitHub are signed with SandFly's key. A build you make yourself is signed with your own debug key, so it can't update a build from GitHub, or be updated by one. Uninstall one before installing the other.
If you build SandFly yourself, you can also build a key into the APK instead. Add gemini.api.key=YOUR_KEY to local.properties in the project folder, then build and install SandFly again. A key entered in settings takes the place of the built-in key. Git ignores local.properties, so your key is not committed, but do not share an APK that contains your key. To use a different model, add gemini.model=MODEL_NAME. The default is gemini-flash-latest.
Library and plugin versions are in gradle/libs.versions.toml. GitHub Actions builds each pull request.
Debug builds include tools to find lag:
- The
BacktalkStalllogcat tag logs each time the main thread is blocked for more than 100 ms, with the code that was running. - The
BacktalkGesturelogcat tag logs touch state changes and each gesture that SandFly detects. - Event processing has trace sections, which show in Perfetto system traces.
- Set
sandflyVersioninversion.gradleto the new version, such as0.2.0, write the release notes inrelease-notes/0.2.0.md, and push to master. List items in the notes start with*: the in-app updater reads lines that start with-as commits. - Tag the commit
v0.2.0and push the tag. GitHub Actions builds it and creates the release, with the changes since the last release.
The build number is the number of commits since buildNumberBase in version.gradle. The version code is 100000000 plus the build number: Gboard leaves typing to the screen reader only when its version code is at least 60105832.
SandFly began as Aaron's fork of Backtalk, Backtalk Aaron, and became a separate project when it had many changes that are not in Backtalk. The git history keeps every change from TalkBack, Backtalk and SandFly. Files from TalkBack and Backtalk are changed in SandFly.
SandFly is under the Apache License 2.0, as TalkBack and Backtalk are. See LICENSE.
The emoji names are from the Unicode CLDR and the Unicode emoji data, copyright Unicode, Inc., used under the Unicode License v3, which is in utils/src/main/assets/emoji_names/LICENSE.
Scripts run in Lua 5.4, copyright Lua.org, PUC-Rio, through LuaJava, copyright gudzpoz, both used under the MIT License. The license texts are in talkback/src/phone/assets/scripting/LICENSE-luajava, which is in the APK.
Japanese braille follows libkuraji, copyright Takuya Nishimoto, ported to Kotlin and used under the BSD 3-Clause License, which is in braille/translate/src/phone/assets/japanese-braille/LICENSE-libkuraji, in the APK. Its dictionary, the JTalk extended dictionary from libkuraji-jtalk-dic (NAIST-JDIC with NVDA Japanese's additions, BSD 3-Clause), is downloaded when Japanese braille is first used, checked against its SHA-256, and read by SandFly's own reader of MeCab's dictionary format.
Dictation recognizes speech on the phone with whisper.cpp 1.9.5, copyright the ggml authors, used under the MIT License, which is in whisper/src/main/cpp/whisper.cpp/LICENSE. Its source is in whisper/src/main/cpp/whisper.cpp, with only the CPU backend. The Whisper models are downloaded from Hugging Face when the user chooses one; they are OpenAI's Whisper models, under the MIT License.
Whisper can also run on Adreno GPUs (Snapdragon phones) with whisper.cpp's OpenCL backend, built as a separate library, libsandfly_whisper_opencl.so (module whisper-opencl, 64-bit ARM only), that is loaded only when Dictation > Processor is GPU. That option appears only on phones with an Adreno GPU and an OpenCL driver for apps. If the library can't load, as with an OpenCL driver older than 3.0, Whisper runs on the CPU. It is built against the Khronos OpenCL headers, Apache License 2.0, in whisper-opencl/src/main/cpp/opencl, and links against a libOpenCL.so made only for linking, which isn't packaged.