Real-time AI audio clarity for rooted Android

A Magisk / KernelSU / APatch module that cleans both your microphone and the audio you hear in real time using open-source speech-enhancement models — configurable from a WebUI.

---

Clarifae installs one native effect library (libclarifae.so) that exposes two descriptors, registered via audio_effects.xml:

• a capture pre-processor on input streams (microphone, calls/VoIP, camcorder, voice recognition) — cleans your outgoing voice; and
• a playback post-processor on output streams (media/music, calls, …) — cleans the audio you hear.

Each side can be toggled independently and is denoised on the fly by one of three engines, selectable as clarity modes.

Note on output + music: these are speech-enhancement models. On voice content (calls, videos, podcasts) output cleanup clarifies nicely; on actual music it treats instruments as noise and degrades them — use the separate Playback strength slider (or turn output off) to taste.

Brand: primary #435787, secondary #68aaf3, font Creative Caps.

Contents

• Clarity modes
• Features
• Requirements
• Quick start
• How it works
• Configuration reference
• clarifae-ctl commands
• Limitations & caveats
• Licensing & credits

Clarity modes

Mode	Engine	Profile	Status
1 · Low latency	RNNoise	Compact RNN denoiser, lowest CPU/latency, 48 kHz framed	✅ Default, works out of the box
2 · Balanced	DTLN	Dual-signal LSTM via TensorFlow Lite, stronger denoising	⚠️ Experimental — needs TFLite + models
3 · Maximum clarity	DeepFilterNet	Full-band 48 kHz deep filtering, highest quality	⚠️ Experimental — needs libdf + model

Modes 2 and 3 are optional builds. If their runtime/models aren't present, Clarifae transparently falls back to RNNoise — it never silently breaks your audio.

Features

• Both directions: mic input and playback output cleanup, each toggleable, from one native AudioFlinger effect (pre-processing + post-processing descriptors).
• Three AI clarity modes with graceful fallback.
• Branded WebUI: master + per-side toggles, mode picker, separate mic/playback strength, max-suppression / voice-gate sliders, high-pass & auto-gain, capture-source and playback-stream selection, live engine status, logs.
• Live, reboot-free settings via persist.clarifae.* properties (the effect re-reads on change).
• clarifae-ctl CLI for scripting and debugging.
• Clean uninstall — the original audio_effects.xml is only shadowed, never edited.

Requirements

• A rooted Android device with Magisk ≥ 24.0, KernelSU, or APatch.
• For the WebUI: KernelSU/APatch have a built-in WebUI launcher; on Magisk, install MMRL to open module WebUIs.
• To build from source: Android NDK (r27+), CMake, and network access for fetch_deps.sh.

Quick start

# from the project folder:
# 1. fetch RNNoise + weights, then build the native effect for all ABIs
./scripts/fetch_deps.sh
./build.sh                      # -> lib/<abi>/

# 2. package a flashable zip
./scripts/package.sh            # -> dist/clarifae-v1.0.0.zip

# 3. flash dist/clarifae-v1.0.0.zip in Magisk/KernelSU/APatch, reboot
# 4. open the Clarifae WebUI, toggle it on, pick a mode

See INSTALL.md for details, optional DTLN/DeepFilterNet builds, and verification steps.

How it works

 INPUT   mic app ─▶ AudioRecord ─▶ capture chain  ─┐  <preprocess> → clarifae_ns
                                                    ├─▶ libclarifae.so (effect.c)
 OUTPUT  media/call ◀─ AudioTrack ◀─ playback mix ─┘  <postprocess> → clarifae_out
                               │  s16/float ⇄ float, live config from persist.clarifae.*
                               ▼
                  backend: rnnoise │ dtln │ deepfilternet
                               │  resample ⇄ engine rate, frame, enhance, wet/dry + atten cap
                               ▼
              cleaned audio ─▶ back to the app (input) / to the speaker (output)

Configuration flows: WebUI → clarifae-ctl → config.conf → persist.clarifae.* → effect. The effect detects a property change (via the property serial) and reloads parameters / re-opens the backend on the next audio buffer, so most changes apply without a reboot. See docs/ARCHITECTURE.md.

Configuration reference

/data/adb/clarifae/config.conf is the source of truth (edited by the WebUI/clarifae-ctl) and is mirrored to system properties the effect reads:

Key / property persist.clarifae.*	Values	Default	Meaning
enabled	0 / 1	1	Master on/off (bypass = passthrough)
in_enabled	0 / 1	1	Clean the microphone (input)
out_enabled	0 / 1	1	Clean playback (output)
mode	1 / 2 / 3	1	RNNoise / DTLN / DeepFilterNet
strength	0–100	70	Wet/dry blend for the mic
out_strength	0–100	60	Wet/dry blend for playback (lower for music)
atten_db	0–60	24	Max suppression depth (protects tone)
vad	0–100	50	Voice-activity sensitivity (gate / auto-gain)
targets	csv of mic,voice_comm,camcorder,voice_rec	mic,voice_comm	Capture sources (input)
out_targets	csv of music,voice_call,ring,alarm,notification,system	music,voice_call	Playback streams (output)
highpass	0 / 1	1	DC/rumble high-pass pre-filter
autogain	0 / 1	0	Gentle make-up gain on speech
loglevel	0–3	1	logcat verbosity (tag Clarifae)

Effect UUIDs: input (pre-proc) c1a71fae-0001-… (type = AOSP noise-suppression); output (post-proc) c1a71fae-0002-… (custom Clarifae type).

clarifae-ctl commands

clarifae-ctl get                 # print config.conf
clarifae-ctl get-json            # settings as JSON
clarifae-ctl set KEY VALUE       # validate + write config + sync property
clarifae-ctl set-many K=V K=V…   # batch set
clarifae-ctl apply               # push all config -> persist.clarifae.*
clarifae-ctl status              # human-readable engine status
clarifae-ctl status-json         # status as JSON (used by the WebUI)
clarifae-ctl restart-audio       # respawn audioserver to reload effects
clarifae-ctl reset               # restore defaults
clarifae-ctl log                 # tail ctl log + logcat -s Clarifae
clarifae-ctl version

Limitations & caveats

• Auto-attach is device-dependent. It edits the device's audio_effects.xml via an overlay. Some vendors use non-standard formats/locations; if patching fails the library is still installed and you can attach it manually (see docs/TROUBLESHOOTING.md).
• Not every app/stream gets cleaned. Effects apply per audio source (input) and per stream type (output). Apps using their own DSP, a raw/unprocessed source, direct/offload output, or exotic routes may bypass it. New sessions pick up changes; some need a restart-audio.
• Output post-processing support varies by device more than input pre-processing — some ROMs ignore or restrict <postprocess> stream effects.
• Music will be degraded if you enable output on the music stream — these are speech models. Use Playback strength or disable playback cleanup for music.
• SELinux/vendor variance. persist.clarifae.* live under default_prop; a defensive sepolicy.rule is included for stricter devices.
• DTLN/DeepFilterNet are experimental and require extra runtimes/models; without them Clarifae uses RNNoise.
• Real-time enhancement adds a small fixed latency (~10 ms for RNNoise + resampler).

Licensing & credits

Clarifae module code is provided as-is for the brand owner. It builds on:

• RNNoise — BSD-3-Clause — https://github.com/xiph/rnnoise
• DeepFilterNet — MIT / Apache-2.0 — https://github.com/Rikorose/DeepFilterNet
• DTLN — see repo — https://github.com/breizhn/DTLN · web: https://github.com/sapphi-red/DTLN-web
• RNNoise voice plugin — https://github.com/werman/noise-suppression-for-voice
• DTLN-AEC (echo cancellation) — https://github.com/breizhn/DTLN-aec
• The vendored native/include/audio_effect.h is a trimmed AOSP-compatible ABI header (Apache-2.0, The Android Open Source Project).

Verify each upstream license before redistribution. TensorFlow Lite (DTLN backend) is Apache-2.0.