SonicLab — AI Audio Studio

A premium, fully offline Android music player & audio toolkit (Kotlin + Jetpack Compose + Media3). A player with real-time DSP effects, an audio toolkit, an analyzer, and on-device AI enhancement — no cloud, no account, no internet.

Table of Contents

• Screenshots
• Offline-first
• Features
• Real-time DSP chain
• Modules (MVVM + Repository)
• Build
• Test, lint & analysis
• Download APK
• Honest status
• Roadmap
• License

Screenshots

This section is waiting for screenshots from a real device. Save the files in docs/screenshots/ named player.png, equalizer.png, settings.png, and studio.png (full details: see docs/screenshots/README.md), then enable the table below:

Offline-first

• The app manifest does not declare the INTERNET permission — nothing "phones home".
• All processing (playback effects, toolkit, analysis, AI models) runs on-device.
• Permissions are limited to features only: reading local audio, foreground media playback, notifications, wake lock, and audio session control.

Features

Library

• Scans local media via MediaStore (with the app's built-in "audio access" permission flow).
• Songs / Albums / Artists tabs: album grid with cover art, artist list, and drill-down into every album/artist.
• Search across songs/artists/albums, favorites, and a track list with cover thumbnails.
• Clear empty states for an empty library, no search results, or missing permission.

Mini-player (all screens)

• Now-playing bar above the bottom navigation: cover, play/pause, next, and a thin progress line; tap to open the full player.

Player

• Media3 ExoPlayer + MediaSessionService: media notifications, lock-screen controls, and Android Auto support; decoder fallback for hi-res/FLAC compatibility.
• Cover header, shuffle, previous / play-pause / next, repeat, and speed presets (0.75×–2×).
• Live pitch (−6..+6 st) — changes the pitch while the track plays, without changing the speed.
• Track format info — codec, sample rate, bit depth, and channels (e.g. FLAC • 96.0 kHz • 24-bit • Stereo) read live from the decoder.
• Queue view — shows only the Up Next tracks (items after the current one); hidden when there are none. Tap to jump, reorder with up/down arrows. From the track list, every track offers Play Next / Add to Queue.
• Crossfade (0–12 s), sleep timer (15/30/60 min), and auto-normalization toward −14 LUFS — real EBU R128 measurement (K-weighting + 400 ms blocks + absolute/relative gating, not plain RMS).
• 3D / 8D mode — presets Off / 3D / 8D / 8D+Center / Surround (quick chips in the player, full controls in the Equalizer). 8D+Center applies 8D rotation to normal audio without mid/side widening so it doesn't sound harsh on headphones, while keeping the center audible. 8D pan never mutes one channel (the Pan Depth slider); Surround adds room echo without rotation. Effects work on every track — including mono and hi-res/FLAC — and toggles apply instantly without changing tracks.

Sound & Effects (Equalizer)

• The Equalizer screen is organized into collapsible sections: EQ Presets, 3D/8D Mode, Tone & Effects. Only the EQ section is expanded on entry; the rest are collapsed by default.
• 3D / 8D Mode — ready-made presets or custom: the 3D, 8D, and Surround switches can be freely combined (mode becomes "Custom"), plus 3D Strength, 8D Rotation Speed, and 8D Pan Depth sliders (limits pan movement so centered sound stays audible on headphones). 8D+Center = 8D on normal audio with a center anchor, no stereo widening.
• Real-time Tone & Effects (always active, audible immediately while playing):
  • Limiter (anti-clip) — final DSP stage: peaks above 0 dB are compressed per-frame (instant attack, smooth release), so full boost presets or loud tracks no longer hard-clip,
  • Treble −12..+12 dB (real-time biquad shelf filter),
  • Reverb (Room) 0–100% + Room Size slider (Schroeder network: 4 combs + 2 all-passes per channel with stereo detune),
  • Pitch (live) −6..+6 st.
• Software 10-band EQ — peaking bands (31.25 Hz–16 kHz, ±15 dB) running inside the DSP chain, consistent across devices (no dependency on the deprecated AudioEffect API), effective even before playback starts. Rendered as a vertical slider visual (boost up, cut down). Bass is a real-time biquad shelf, plus stereo Balance and presets.
• Reset buttons (all / per-section) ask for confirmation before resetting.
• Presets: Music HD is smoother, Gaming (V-shape + virtualizer) differs from Car Audio (strong bass + boosted mids), BT Speaker focuses on mid-bass 100–250 Hz, plus Bass + Vocal, Acoustic, and Jazz (warm room reverb).
• Automatic persistence — balance, 3D/8D mode, treble, reverb, pitch, bass, and EQ bands are restored when the app reopens.

Studio (single screen — opened from the "More" menu of any library track)

• The selected library track automatically becomes File 1 (falls back to the currently playing track when opened without a selection), with analyzer + toolkit in one place:
  • File Info — decode/codec info,
  • Convert to WAV — transcode to WAV via MediaCodec,
  • Cut — export the first 30 seconds.
• DSP engine (on-device) — fully wired to the UI (full-file processing is capped at ~10 minutes per file for memory safety):
  • Join — merge two files into one WAV,
  • Reverse — play the audio backwards,
  • Pitch (−12..+12 st) and Tempo (0.5×–2×) — interactive WSOLA sliders,
  • Normalizer (−14 LUFS) — EBU R128-based loudness normalization,
  • Vocal Remover — instrumental output (requires a stereo file; STFT center-channel with soft ratio mask + smoothing).
• Tool results — every finished tool shows a Result Ready card with Play (play the result immediately), Share (share sheet via FileProvider), and Save to Downloads (MediaStore, Android 10+). Status messages (success/failure) appear as a dismissible Snackbar.
• Analyzer — one-tap analysis for the playing track or a picked file: waveform preview, audio file info, and EBU R128 LUFS loudness via the PcmReader pipeline.

AI (on-device)

• AI Enhance — real-time enhancement on the playback path via Media3 AudioProcessor, driven by NeuralEnhancer: it loads a bundled TFLite model when present and otherwise falls back to the transparent DSP enhancer (adaptive gain toward −20 dBFS + soft-knee limiter, no EQ coloring, no hard clip). Works on every track, including hi-res/FLAC. The Settings screen reports which engine is actually active.
• Vocal separator — STFT center-channel with soft ratio mask (no neural model).
• Native low-latency engine — the dsp module ships a real Oboe (C++) engine compiled to .so (libsoniclab_oboe.so, 4 ABIs): 48 kHz float capture, playback, and full-duplex on the fast audio path, wired to Kotlin through JNI (OboeNativeEngine). It is not yet connected to the Media3 sink; that is on the roadmap.
• No neural models are bundled: the previous denoiser_v1.tflite was trained on synthetic 16 kHz audio and degraded real music into buzzing artifacts, so it was removed in favor of the transparent DSP enhancer. The strongest open neural enhancer (DeepFilterNet, Apache-2.0/MIT) is ONNX-only, which needs ONNX Runtime Mobile rather than TFLite — see On-device AI models.

Settings

• AMOLED theme (pure black), AI Enhance toggle (real-time), Auto Normalization (EBU R128), Direct Mode — rebuilds the sink without the DSP chain for the cleanest output path, Crossfade slider, Sleep Timer, AI model status, and About App & Developer (full info screen).
• Output Audio (output path quality):
  • Hi-Res 24-bit Output — with the DSP chain active the whole chain runs in 32-bit float through a custom AudioSink wrapper and reaches AudioTrack as float PCM (never down-converted), so effects and hi-res work together; the toggle shown in Direct mode controls media3's own float path for the cleanest bypass. The player rebuilds automatically while preserving the queue & position,
  • Headroom −3..0 dB — headroom before effects so EQ/presets are less likely to clip.

UI & icons

• Consistent Indonesian UI throughout the app.
• The currently playing track is highlighted in the track list — colored title + player icon.
• Premium dark theme (purple/cyan palette) with an AMOLED option.
• Premium launcher icon: adaptive icon (gradient + waveform EQ with glow), round icon, and monochrome for Android 13+ themed icons.
• Empty states, screen transitions (fade + slide), About screen, haptic feedback on main controls, and album covers downsampled for RAM efficiency.

Real-time DSP chain

All playback effects run in one Media3 AudioProcessor chain inside DefaultAudioSink:

balance → headroom → gain (auto-normalization) → 10-band EQ → tone (bass/treble) → reverb (room) → AI enhance → spatial (3D/8D) → limiter (anti-clip)

Direct mode bypasses the whole chain (the service rebuilds the player without AudioProcessors; position and queue are preserved).

All processors extend PcmAudioProcessor, which:

• accepts PCM16 and PCM float (never fails on hi-res/FLAC tracks),
• stays always active so effect toggles apply instantly without seeking/track changes,
• up-mixes mono → stereo for 3D/8D effects,
• passes audio through untouched (zero-copy) when the effect is off,
• never re-encodes to PCM16 during playback: the chain stays in 32-bit float end-to-end (the 16-bit converter is only a format fallback),
• runs inside DspAudioSink, which normalizes every renderer buffer (16/24/32-bit or float) to 32-bit float before the chain and feeds the output to a float-capable DefaultAudioSink — hi-res quality is preserved end-to-end while the DSP chain stays active.

Modules (MVVM + Repository)

Modular structure (4 modules) with clear separation of concerns: data, audio processing, playback, and UI. Code stays grouped by package (com.soniclab.*), so the logical layers remain visible.

• :core — domain models (Track/Playlist/Preset), result wrapper, permissions, time utils, MediaStore scanning, playlists/favorites, DataStore preferences
• :dsp — audio signal processing & analysis: audioengine (Oboe scaffold), DSP toolkit & WAV conversion (MediaCodec), analyzer (FFT, R128 LUFS, waveform, codec info), and on-device AI (TensorFlow Lite + DSP fallback)
• :player — Media3 ExoPlayer + MediaSessionService (notifications, lock screen, Android Auto), PcmAudioProcessor chain (balance/headroom/gain/10-band EQ/tone/reverb/enhance/spatial/limiter), Direct mode, hi-res float output, speed, crossfade, sleep timer, live pitch via PlaybackParameters
• :app — Compose UI (theme, navigation, screens, About) + spectrum/waveform visualizer

Build

Requirements: JDK 17+, Android SDK with platforms;android-36 and build-tools;36.0.0.

./gradlew :app:assembleDebug

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

Test, lint & analysis

./gradlew test               # all JVM unit tests
./gradlew :dsp:testDebugUnitTest      # DSP/PCM toolkit unit tests
./gradlew :player:testDebugUnitTest   # player DSP chain unit tests (balance/headroom/gain/eq/tone/reverb/spatial/limiter)
./gradlew :app:lintDebug     # Android Lint

Status in this repo: build passes, unit tests pass (13/13 in :dsp, 23/23 in :player), lint 0 errors. CI runs the same checks on every push.

Download APK

This repo is public and GitHub Actions builds a debug APK on every push to main:

1. Open Actions → latest green Android Build run.
2. Download the soniclab-debug-apk artifact.
3. Install app-debug.apk on a device (Android 8.0+ / API 26+).

Or build locally with the command above.

On-device AI models

The app currently ships without neural models: the synthetic denoiser was removed because its quality on real music was below the transparent DSP enhancer. The best open neural enhancer available today is DeepFilterNet (github.com/Rikorose/DeepFilterNet) — Apache-2.0/MIT, ~4.6k stars, trained for speech denoising at 48 kHz. Its models are ONNX, so a real integration would add ONNX Runtime Mobile; forcing them into TFLite is fragile and loses quality. And because DeepFilterNet targets speech, applying it to full-band music audibly changes the timbre — bundling it now would repeat the denoiser_v1 mistake.

That is why the AI Enhance path is a seam: NeuralEnhancer activates automatically the moment a contract-compliant .tflite model appears at models/ai_enhancer_v1.tflite (see dsp/src/main/assets/models/README.md), and otherwise stays on the transparent DSP enhancer. A model trained on real music is on the roadmap. Training scripts, when reintroduced, will live in scripts/ (pure NumPy + flatbuffers, no TensorFlow/PyTorch toolchain).

Honest status

• Verified by CI: compile + unit tests + lint green; debug APK generated automatically.
• Not yet verified: runtime behavior on real devices (Direct mode & MediaSession reconnect, MediaSession/notifications, AI Enhance quality on real music, toolkit processing of large files, real-time DSP effect results). This needs manual testing on a physical Android device.

Roadmap

• Wire OboeNativeEngine into the playback sink (custom AudioSink on the fast audio path, replacing the current Media3 DefaultAudioSink wrapper) so the DSP chain runs inside the native engine; add live mic monitoring via the full-duplex stream.
• Higher-quality AI model: either a music-trained TFLite model dropped into assets/models/, or DeepFilterNet via ONNX Runtime Mobile.
• More WSOLA quality knobs (quality presets in the UI).
• Migrating UI strings to resources (strings.xml) for formal multi-language support.
• Signed release build.

License

Apache License 2.0 © 2026 soe1hom-arch.

Every source file in this repository carries an SPDX header (SPDX-License-Identifier: Apache-2.0) so the license is machine-readable and traceable per file.

Third-party components (AndroidX, Media3, Oboe, TensorFlow Lite, Kotlin, …) are credited in THIRD_PARTY_NOTICES.md; the APK also bundles the notices and shows them in Tentang → Kredit pihak ketiga.