WELCOME to WakkaQt v2.3

–– um app multimídia studio de karaoke, multiplataforma e livre

click the logo to download Windows x64 ZIP bundle

Free as in Freedom!

WakkaQt 🎤

“Because Auto-Tune is expensive and your bathroom acoustics only get you so far.”

WakkaQt is a free, open-source karaoke recording and production studio built with Qt6. Load a karaoke video, grab a mic, sing your heart out, and walk away with a finished, mixed, pitch-corrected MP4 — complete with a webcam feed, vocal overlay, and a pitch indicator that will mercilessly show the world every flat note you tried to sneak past.

No subscriptions. No cloud. No judgment. (Well, maybe a little judgment from the pitch monitor.)

Current version: 2.3

Downloads

Windows binaries — ready to run, no setup required: 👉 https://gu.pro.br/WakkaQt

Linux users — you get to build it yourself, which is a feature, not a bug. See build instructions below.

What It Does

1. Plays Your Karaoke Track

Drop in any MP4, MKV, WebM, AVI, MOV, MP3, WAV, FLAC, or OPUS file. If Qt6 Multimedia can play it, WakkaQt will play it. You can also paste a YouTube URL and download the video directly from inside the app (powered by yt-dlp).

2. Records You Singing It

Select your microphone from a list of all detected devices. Hit 🎤 SING. Optionally capture your webcam at the same time — for those who want to remember exactly what they looked like belting out Bohemian Rhapsody at 2 AM.

3. Makes You Sound Better Than You Are

Before rendering, the vocal track runs through a full DSP pipeline:

• Pitch correction — phase-vocoder pitch shifting with adjustable strength (0 = raw humanity, 100 = robot perfection)
• Scale-aware snapping — snap pitch to Major, Minor, Pentatonic, Blues, Dorian, Mixolydian, Lydian, Phrygian, Locrian, Harmonic Minor, Melodic Minor, Whole Tone, Diminished, or plain Chromatic — in any of the 12 keys
• Retune speed — 0 ms for that T-Pain effect, up to 300 ms for a natural glide
• Formant preservation — LPC-based envelope re-synthesis keeps your voice sounding human even after aggressive pitch shifting
• Noise reduction — spectral subtraction gate with adaptive noise floor estimation (goodbye, fan noise)
• Reverb — Freeverb-style Schroeder reverb with room size, decay, and wet/dry controls
• Dynamics — compressor, limiter, and harmonic exciter for a polished, loud-enough final mix

All FFTW plans are created once and reused for the entire recording — no plan allocation mid-session, no glitches.

4. Lets You Preview and Tweak

A full-featured preview dialog lets you hear the processed vocal, adjust every enhancement parameter in real time, nudge the audio/video sync offset, and preview again — as many times as you need before committing to a render.

5. Renders a Professional-Looking Video

Output: a 1920×1080 MP4 with the karaoke video on top and your webcam below. The vocal is mixed in with all enhancements applied. A pitch indicator strip is burned into the webcam frame — green when you’re in tune, yellow when you’re drifting, red when you’re… having a moment.

Native FFmpeg integration renders entirely in-process with a real-time progress bar. Falls back gracefully to spawning ffmpeg via a subprocess if the dev libraries weren’t present at build time.

6. Generates a Backing Track

Load any song and click 🎵 Backing Track. WakkaQt downloads the UVR-MDX-NET-Inst_HQ_3 ONNX vocal separation model (~80 MB, once) and runs it locally on your machine — no internet required after the first download, no cloud service, no privacy leak, no subscription.

The model separates vocals from the instrumental using MDX-Net deep learning, processed through a full STFT/iSTFT pipeline with FFTW3. If the input is a video file, the separated audio is muxed back onto the original video by default — so you keep the visuals. You can still save as audio-only (WAV or MP3) by choosing an audio format in the save dialog. Separation can be aborted mid-way with the Abort button.

7. Keeps a Session Library

Every recording is saved to ~/.WakkaQt/library/ with a UUID folder, all source files, and JSON metadata. The library dialog lists everything with timestamps. You can rename, delete, or re-render any session — with updated enhancement settings — at any time.

Feature Overview

Changelog

v2.3 — YouTube karaoke browser & hardware-accelerated rendering

• YouTube karaoke browser — new “🔍 Browse YouTube” button opens a search dialog with two scrollable rows of cards (karaoke results and originals). Thumbnails load asynchronously via QNetworkAccessManager. Clicking a card shows a preview panel with a larger thumbnail, title, channel, and duration; clicking “Download this video” fills the URL field and launches the existing yt-dlp download flow. Powered by yt-dlp --flat-playlist --dump-json
• Hardware-accelerated H.264 encoding (VAAPI / NVENC) — renderVideo now probes available hardware encoders at render time in order: h264_nvenc → h264_vaapi → h264_v4l2m2m, falling back to software libx264. A dummy frame is encoded to confirm the driver actually works before committing (needed because avcodec_open2 succeeds even when hardware encoding will silently fail). For Intel VAAPI, a fast YUV420P→NV12 conversion is applied per frame so the software pipeline stays in YUV420P while the VAAPI surface gets the NV12 format Intel iHD requires
• Dialog close properly cancels background work — closing the backing-track conversion dialog via the window X button now sets the cancellation flag, aborting the ONNX separation thread. Previously the thread ran to completion invisibly and could show a phantom save dialog. Same fix applied to DownloadDialog (yt-dlp processes are now killed on X via closeEvent) and PreviewDialog (in-flight enhancement futures and ffmpeg processes are cancelled before hiding)
• Fixed -Wstringop-overflow false positives in vocalenhancer.cpp — GCC’s false positive on QVector(N, 0.0) silenced with early-exit guards in applyLPCInverseFilter and applyLPCSynthesisFilter

v2.2.3 — Video recording/render improvements

• A/V interleaving fixed — audio and video packets are now written interleaved by DTS in both the karaoke render and the backing-track mux. Previously all audio was written before any video, causing players to show no video from the start and no audio after seeking
• Raw camera input preferred — the camera format selector now picks raw/uncompressed formats (NV12, YUYV, etc.) over JPEG when available, giving the recording encoder pristine sensor data instead of pre-compressed frames. JPEG remains the fallback
• H264 preferred over MotionJPEG for recording — H264 delivers far better quality at the same bitrate; MotionJPEG dropped to fallback in the codec preference order
• H265 moved to last in recording codec preference — H265 in Matroska is unreliable in Qt multimedia on Linux and caused a hard crash; it remains available as a last resort only
• Fixed segfault on recording error — destroying the QMediaRecorder object from inside its own errorOccurred signal handler was undefined behaviour. The media reset is now deferred via QTimer::singleShot so it runs after the signal unwinds
• Higher render quality — output bitrate raised from 3 Mbps to 5 Mbps, H264 encoder preset changed from fast to medium, and frame scaling upgraded from bilinear to bicubic

v2.2.1

• Abort render / abort separation — both the video render and the vocal separation now have an Abort button that cancels the operation mid-flight without leaving corrupted output files
• Video-preserving backing track output — when the input to the vocal separator is a video file, the separated instrumental is muxed back onto the original video by default, preserving the visuals. Choosing an audio format (WAV, MP3) in the save dialog still saves audio only
• Extension auto-follows filter in all save dialogs — switching format in any save dialog (render, library export, backing track) now updates the output filename extension automatically. Achieved via QFileDialog::DontUseNativeDialog + filterSelected signal across all three dialogs
• Native FFmpeg for all separator I/O — decodeToFloatStereo, writeFloatWav, transcodeAudio, and muxVideoWithAudio are now implemented natively in-process via libavformat/libavcodec/libswresample. The old QProcess ffmpeg / QProcess ffprobe calls in the separator remain as #else fallbacks behind #ifdef WAKKAQT_FFMPEG_NATIVE guards
• Pitch monitor height reduced — PitchMonitorWidget maximum height lowered from 48 px to 24 px for a more compact layout

v2.1.3

• Generate Backing Track — new feature powered by the UVR-MDX-NET-Inst_HQ_3 ONNX model. Downloads the model on first use (~80 MB), then separates vocals from any loaded track entirely offline. Output can be saved as WAV or MP3. The button appears automatically when a file is loaded and disappears when it isn’t needed
• Save to existing file fixed — overwriting an already-existing WAV with the backing track output no longer fails silently and loses the processed audio. The destination is removed first; on copy failure the temp file is preserved and its path is shown in the error dialog
• Playback stopped before separation — WakkaQt now stops playback before starting the ONNX separation process to free audio and CPU resources for the model run
• CMake: ONNX Runtime optional dependency — ONNX Runtime is detected at configure time; if absent, the backing-track feature is disabled gracefully with a build warning and everything else continues to work

v2.0.2

• Pitch overlay on rendered video — note name (e.g. “C#4”), cents deviation bar, and “+/−Nc” value burned into the bottom of the webcam track. Color-coded: green < 10 ¢, yellow < 30 ¢, red ≥ 30 ¢
• PitchMonitorWidget always visible — pitch display is shown at all times, not just during recording
• Library: multi-select deletion — Ctrl+click or Shift+click to select multiple sessions for batch delete
• Native sample rate preservation — recorded audio is processed at the microphone’s native sample rate (e.g. 48 kHz); no forced downsampling to 44100 Hz
• Library: tuned.wav not saved — the intermediate tuned vocal is regenerated from the raw audio on every re-render, so the preview dialog always reflects current settings
• A/V sync: fixed video desync with manual offset — both audio and video offsets are now set to manualOffset directly; previously a formula involving system latency caused the video to lag or lead by hundreds of milliseconds

v2.0.1

• VocalEnhancer: fixed progressive robotic distortion — compressor and harmonic exciter were running inside the pitch-correction loop; now applied exactly once after all pitch work
• VocalEnhancer: fixed stale pitch state at section boundaries — pitch state is cleared after ~400 ms of silence, eliminating wrong shift ratios at phrase starts
• VocalEnhancer: octave detection guard — 2× / 0.5× detection errors corrected when pitch confidence is low
• A/V sync: fixed negative manual-offset case — webcam pre-roll footage is correctly seeked past in both directions
• Rendering: UI fully locked during FFmpeg — all controls disabled for the entire render regardless of code path
• FFmpeg native: keyframe alignment correction — PTS of the first decoded frame is used to shift all subsequent video frames into exact alignment after a seek

Building on Linux

Prerequisites

Install the required development packages (Debian/Ubuntu):

sudo apt install \
    build-essential cmake ninja-build \
    qt6-base-dev qt6-multimedia-dev \
    libqt6multimedia6 libqt6multimediawidgets6 \
    libfftw3-dev \
    libavformat-dev libavcodec-dev libavfilter-dev \
    libavutil-dev libswresample-dev libswscale-dev \
    libglib2.0-dev \
    pkg-config

H264 recording support (strongly recommended)

The standard libavcodec package on Ubuntu/Debian is built without H264 encoding. Install the extra-codecs variant so WakkaQt can record webcam footage in H264 (the preferred codec):

sudo apt install libavcodec-extra

This replaces libavcodec with a version that has H264 (and other patent-encumbered codecs) compiled in. Without it, WakkaQt falls back to MotionJPEG for recording, which gives lower quality at the same bitrate.

On Fedora / RHEL, the ffmpeg-free package in the standard repos also lacks H264 encoding. Enable RPM Fusion and swap in the full FFmpeg build:

sudo dnf install https://mirrors.rpmfusion.org/free/fedora/rpmfusion-free-release-$(rpm -E %fedora).noarch.rpm
sudo dnf swap ffmpeg-free ffmpeg --allowerasing

For the AI backing-track feature, also install the ONNX Runtime development package.

Debian/Ubuntu:

sudo apt install libonnxruntime-dev

Fedora / RHEL (not in standard repos — install from the official release):

ORT_VERSION=1.20.1
wget https://github.com/microsoft/onnxruntime/releases/download/v${ORT_VERSION}/onnxruntime-linux-x64-${ORT_VERSION}.tgz
tar -xzf onnxruntime-linux-x64-${ORT_VERSION}.tgz
sudo cp -r onnxruntime-linux-x64-${ORT_VERSION}/include/onnxruntime /usr/local/include/
sudo cp onnxruntime-linux-x64-${ORT_VERSION}/lib/libonnxruntime.so* /usr/local/lib/
sudo ldconfig

Windows (MinGW / MSVC):

1. Download onnxruntime-win-x64-*.zip from the ONNX Runtime releases page.
2. Extract and copy the contents to C:\Program Files (x86)\onnxruntime\.
3. Inside that folder, create the subfolder include\onnxruntime\ and move all header files from include\ into it. The final structure must be:C:\Program Files (x86)\onnxruntime\ include\ onnxruntime\ onnxruntime_cxx_api.h onnxruntime_c_api.h ... (all other .h files) lib\ onnxruntime.lib onnxruntime.dll
4. Copy onnxruntime.dll next to WakkaQt.exe in your build/install folder — Windows needs the DLL at runtime.

CMake will find the library automatically in Program Files (x86)\onnxruntime during configure.

ONNX Runtime is entirely optional — if it is not found at configure time, WakkaQt builds and runs normally without it and the backing-track button simply won’t appear.

On Fedora/RHEL-based systems:

sudo dnf install \
    cmake ninja-build gcc-c++ \
    qt6-qtbase-devel qt6-qtmultimedia-devel \
    fftw-devel \
    ffmpeg-free-devel \
    glib2-devel \
    pkgconf

You also need ffmpeg and yt-dlp installed as runtime tools:

sudo apt install ffmpeg yt-dlp   # Debian/Ubuntu
sudo dnf install ffmpeg yt-dlp   # Fedora

Configure and Build

git clone https://github.com/guprobr/WakkaQt.git
cd WakkaQt
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel

Debug build:

cmake -B build -DCMAKE_BUILD_TYPE=Debug
cmake --build build --parallel

Run

./build/WakkaQt

Install System-Wide (optional)

sudo cmake --install build

Installs to /usr/bin/WakkaQt, with an icon at /usr/share/icons/hicolor/256x256/apps/WakkaQt.png and a .desktop launcher in /usr/share/applications/.

Runtime Dependencies

Both must be on $PATH at runtime. The ONNX model (~80 MB) is downloaded automatically on first use of the backing-track feature and cached in ~/.WakkaQt/models/.

License

See LICENSE for details.