HachiTune

A lightweight pitch editor built with JUCE, featuring neural pitch detection and vocoder resynthesis for real-time preview.

Overview

HachiTune is a standalone app and audio plugin for editing vocal pitch in a piano roll. It analyzes audio using neural pitch detectors (RMVPE / FCPE), segments notes with the GAME model, and resynthesizes edits in real time through a neural vocoder (PC-NSF-HiFiGAN). Edits are previewed instantly with minimal latency and exported as audio or MIDI. The plugin supports both ARA hosts and a non-ARA capture workflow.

Tips

Due to the project's complexity and our ongoing efforts to standardize and decouple the codebase, we are currently carrying out a partial refactor, including a completely new GUI interface. As a result, the project's progress may be slightly delayed. Thank you for your understanding.

Features

• Piano roll editing with select, draw, split, and stretch modes
• Note timing stretch with absorb (zero-sum) and ripple (shift subsequent) modes
• Neural pitch detection (RMVPE default, FCPE optional)
• GAME-based note segmentation with D3PM diffusion model
• Pitch curve editing with per-note delta, scale, offset, tilt, and variance controls
• Real-time incremental vocoder resynthesis (PC-NSF-HiFiGAN)
• Undo/redo for all editing operations
• Export audio (WAV) and MIDI
• Standalone app plus VST3/AU plugins (AAX optional, ARA supported)
• GPU acceleration via CUDA or DirectML on Windows; CoreML on macOS
• Multi-language UI (English, Japanese, Chinese Simplified, Chinese Traditional)

Download

To Release

Requirements

• CMake 3.22+
• C++17 compiler
• Git (for submodules)

Platform notes

Platform	Requirements
Windows	Visual Studio 2019+ with C++ workload, Windows SDK
macOS	Xcode + command line tools
Linux	GCC/Clang, ALSA development libraries

Models and Assets

The build expects required model files in Resources/models and will fail if any are missing:

• pc_nsf_hifigan.onnx (vocoder)
• rmvpe.onnx (pitch detection)
• fcpe.onnx (pitch detection, optional)
• GAME/ (note segmentation: encoder.onnx, estimator.onnx, segmenter.onnx, dur2bd.onnx, bd2dur.onnx, config.json)
• cent_table.bin
• mel_filterbank.bin

Other runtime assets live in Resources/lang and Resources/fonts.

Build

Submodules

git submodule update --init --recursive

Configure

cmake -B build -DCMAKE_BUILD_TYPE=Release

Build

cmake --build build --config Release

Optional flags

• -DUSE_CUDA=ON Enable CUDA execution provider (Windows)
• -DUSE_DIRECTML=ON Enable DirectML execution provider (Windows)
• -DUSE_ASIO=ON Enable ASIO support (Windows)
• -DUSE_BUNDLED_CUDA_RUNTIME=ON Bundle CUDA redistributables (Windows)
• -DUSE_BUNDLED_DIRECTML_RUNTIME=ON Bundle DirectML runtime (Windows)
• -DONNXRUNTIME_VERSION=1.24.3 Override ONNX Runtime version

Notes:

• CUDA and DirectML cannot be enabled at the same time.
• ONNX Runtime is downloaded during CMake configure.

Plugin Formats

• VST3 and AU are built by default.
• AAX is built only when AAX_SDK_PATH is set and valid.
• ARA is enabled when the third_party/ARA_SDK submodule is present.

Usage

1. Open an audio file (WAV/MP3/FLAC/OGG).
2. Analyze to detect pitch and segment notes.
3. Edit notes in the piano roll — adjust pitch, timing, and expression.
4. Use stretch mode to fine-tune note boundaries (absorb or ripple).
5. Preview changes in real time via the neural vocoder.
6. Export the edited audio or MIDI.

Plugin modes

• ARA mode: Direct audio access in supported hosts (e.g., Studio One, Cubase, Logic).
• Non-ARA mode: Auto-capture and process in hosts without ARA.

Project Structure

HachiTune/
  Source/
    Audio/        # Audio engine, pitch detection, synthesis, vocoder
    Models/       # Project, note, and audio data models
    UI/           # Piano roll, toolbar, and UI components
    Utils/        # Resampling, localization, mel spectrogram utilities
    Undo/         # Undo/redo action system
    Plugin/       # VST3/AU/AAX/ARA plugin integration
  Resources/
    models/       # ONNX models and data files
    lang/         # Localization JSON (en, ja, zh, zh-TW)
    fonts/        # UI fonts
    svg/          # Toolbar icons
  third_party/
    JUCE/         # JUCE framework (submodule)
    ARA_SDK/      # ARA SDK (submodule, optional)
  CMakeLists.txt

Acknowledgements

This project uses the following models from the open-source community:

• GAME — Note segmentation model by OpenVPI
• PC-NSF-HiFiGAN — Neural vocoder by OpenVPI
• RMVPE — Robust pitch estimation model by yxlllc

License

See LICENSE.