Voice Isolation for Android

Android app that captures regular media and game audio through Shizuku and processes it with an on-device streaming ML model, with independent background reduction and voice boost.

What works

• One tap on START connects to the already running Shizuku service, loads GTCRN, and configures the audio route.
• A Quick Settings tile starts or stops isolation without opening the app; long-pressing the tile opens the app.
• A persistent notification shows the active service and provides intensity and stop actions.
• Reduce background controls attenuation of the estimated background from 0–100%; Boost voice independently raises the estimated voice level from 0 to about +6 dB. At 0% reduction and 0 dB boost, the latency-aligned original is preserved.
• Optional Auto mode uses GTCRN output energy relative to the aligned original to estimate speech. It holds through short pauses and smoothly softens background reduction between dialogs. This is a speech estimate, not a separate speech-recognition model; vocals and speech-like instruments can trigger it.
• Adaptive playback increases the active buffer when new underruns occur, within a preallocated limit. The UI displays underruns, actual buffer duration, and Normal / Buffered / Safe status. A larger buffer can absorb scheduling jitter but adds latency and cannot fix sustained processing slower than real time.
• Podcast, Film, Gaming, Maximum, Custom, and named user presets persist across app restarts and appear in the notification and Quick Settings tile. Named presets now include both controls and Auto mode. Existing presets retain their previous sound; built-in presets keep Auto mode off.
• An auto-off slider can stop isolation after 5–60 minutes without captured audio, or the feature can be disabled.
• An opt-in, token-protected automation API supports start, stop, preset, and intensity commands from Tasker or ADB.
• The Maximum preset compensates the model's measured 2.3 dB voice attenuation while retaining background suppression. The limiter protects output peaks when boosting voice.
• Processing stays entirely on-device; the app intentionally has no internet permission.
• A privileged AOSP AudioPolicy loopback route replaces the original mix without changing the global media volume.
• Stopping releases AudioRecord, AudioTrack, ONNX Runtime, and the dynamic audio policy.

Requirements

1. Android 13 or newer.
2. Install Shizuku.
3. Start Shizuku. Without root, this is required again after every reboot; Android 11+ can use wireless debugging.
4. Install the app and grant Shizuku access on first use.
5. Start a media app and tap START in Voice Isolation.
6. Optional: add Voice Isolation from the Quick Settings edit screen for one-tap access.

Once Shizuku is running and permission has been granted, starting the app is fully automatic. Without root, the app cannot start Shizuku by itself after a device reboot.

Tasker and automation

Enable Automation API in the app and copy its per-install token. The exported receiver has no intent filter; callers must target it explicitly and include that token. Keep the token private.

Actions:

Action	Extra
dev.sk2andy.voiceisolation.automation.START	token
dev.sk2andy.voiceisolation.automation.STOP	token
dev.sk2andy.voiceisolation.automation.SET_PRESET	token, preset: podcast, film, gaming, maximum, custom, or a saved preset name
dev.sk2andy.voiceisolation.automation.SET_INTENSITY	token, intensity: number from 0.0 to 1.0

Configure Tasker's Send Intent action with target Broadcast Receiver, package dev.sk2andy.voiceisolation, class dev.sk2andy.voiceisolation.service.AutomationReceiver, and one action from the table. Example ADB commands:

adb shell am broadcast \
  -n dev.sk2andy.voiceisolation/.service.AutomationReceiver \
  -a dev.sk2andy.voiceisolation.automation.START \
  --es token YOUR_TOKEN

adb shell am broadcast \
  -n dev.sk2andy.voiceisolation/.service.AutomationReceiver \
  -a dev.sk2andy.voiceisolation.automation.SET_PRESET \
  --es token YOUR_TOKEN \
  --es preset film

adb shell am broadcast \
  -n dev.sk2andy.voiceisolation/.service.AutomationReceiver \
  -a dev.sk2andy.voiceisolation.automation.SET_INTENSITY \
  --es token YOUR_TOKEN \
  --ef intensity 0.8

Android may reject automatic starts while the app is background-restricted. In that case, exempt Voice Isolation from battery restrictions or trigger start from a user-initiated Tasker action. Shizuku must already be running and authorized.

SET_INTENSITY changes background reduction while preserving voice boost and Auto mode. SET_PRESET restores the complete saved mix; built-in presets restore their original tuning with Auto mode off.

Languages

English is the default. The app also includes Spanish, Swedish, Polish, French, Portuguese, Russian, and Simplified Chinese translations. Android selects the language from the device or per-app language settings.

Technical limitations

Area	Behavior
Regular media and game audio	Processes USAGE_MEDIA, USAGE_GAME, and USAGE_UNKNOWN
Calls, alarms, notifications	Intentionally not captured
DRM, hardware offload, tunneled playback	May bypass the framework mix route and remain unchanged
Vendor ROMs	Dynamic shell AudioPolicy is AOSP-based but not guaranteed on every OEM device
Model quality	GTCRN performs speech enhancement, not perfect vocal separation; music is usually attenuated strongly, but complete removal is not guaranteed
Shizuku through ADB	Must be restarted after every device reboot

Architecture

Media apps
    │
    ▼
AOSP AudioPolicy: pure LOOP_BACK (Shizuku UID shell/root)
    │  16 kHz mono PCM
    ▼
AudioRecord ─┬──────────────────── delayed dry ────────────────┐
             └→ STFT → GTCRN/ONNX Runtime → iSTFT → +2.3 dB ─┤
                                                                ▼
                                                     intensity mix → AudioTrack

The app excludes its own AudioTrack from the capture mix by Shizuku UID. Android 13+ also receives the correct com.android.shell attribution so AudioFlinger accepts the shell identity. On Android 12, the ADB shell lacks the required MODIFY_AUDIO_ROUTING permission, so that version is not listed as compatible.

Build

./gradlew assembleDebug

APK: app/build/outputs/apk/debug/app-debug.apk

Tests:

./gradlew testDebugUnitTest
./gradlew connectedDebugAndroidTest
./gradlew lintDebug

Verification

Test	Coverage
JVM unit tests	PCM boundaries and STFT/iSTFT reconstruction
Android instrumentation	Real ONNX Runtime and gtcrn.onnx loading, followed by four streaming frames
Android 16 emulator	Shizuku UserService as UID 2000, policy registration, remote submix, 440 Hz USAGE_MEDIA capture, GTCRN processing, and stop/policy unregister
Speech/music vector	Pinned real GTCRN speech sample plus a deterministic music bed; requires at least 2 dB SI-SDR improvement and 6 dB speech-over-music selectivity

On the emulator, the synthetic signal took about 0.9 ms of inference time per 16 ms frame (RTF ≈ 0.05). Actual performance depends on the device.

The reproducible model measurement runs separately so the Android app itself can remain offline:

python3 -m venv /tmp/voice-isolation-validation
/tmp/voice-isolation-validation/bin/pip install numpy onnxruntime
/tmp/voice-isolation-validation/bin/python tools/validate_gtcrn_music.py

The pinned vector measured +2.81 dB SI-SDR improvement, -2.30 dB speech gain, and -16.11 dB music gain. Speech therefore remained 13.80 dB above the music gain. This demonstrates strong attenuation for this test vector, not perfect separation for arbitrary music.

Acknowledgements and third-party software

Poise Android was used only as a product and architecture comparison. That project does not publish a license, so no Poise source code was copied. The GTCRN model comes directly from the MIT-licensed official upstream project. See THIRD_PARTY_NOTICES.md for details.

The app also uses ONNX Runtime and the Shizuku API. Thanks to their maintainers and contributors.

License

Original source code is available under the MIT License. Third-party components remain under their respective licenses; see THIRD_PARTY_NOTICES.md.