v2.0.78-forge.0.1.1

VCClient Forge v2.0.78-forge.0.1.1

This release introduces a client-side TCP proxy mode (--remote) for remote backend setups, native startup error dialogs in the desktop application, and fixes session state persistence so that Passthrough and Recording always initialize in a clean default state.

---

Highlights & What’s New

🌐 Remote Backend TCP Proxy (--remote <host:port>)

• Remote WebAudio Without Certificates: Modern browsers and embedded webviews require a Secure Context (https:// or 127.0.0.1 / localhost) to enable microphone access. When connecting directly to a remote server over LAN via HTTP, microphone capture is blocked by default.
• Loopback Forwarding: Launching the desktop application with --remote <host:port> (e.g., vcclient-forge.exe --remote 192.168.1.50:18888) automatically establishes a local TCP proxy listening on 127.0.0.1:18888. Because browsers treat loopback origins as trustworthy, microphone access (getUserMedia) and WebAudio processing work out-of-the-box without generating self-signed certificates or handling browser SSL warning screens.

🛡️ Native Error Dialogs & Robust Process Lifecycle

• Startup Error Visibility: If the backend process fails to start (e.g., port conflict, missing dependencies, or crash), the desktop application now presents a native dialog with the error message before closing, rather than exiting silently.
• Reliable Shutdown: Ensures all child processes and background proxy listeners terminate cleanly whenever the application is closed.

🔧 Clean Startup State (Passthrough & Recording)

• Session State Reset: Fixed an issue where closing the application with Passthrough or Recording enabled caused subsequent launches to automatically start with Passthrough active.
• Default State Initialization: Transient runtime toggles are now excluded from config persistence on disk, ensuring the audio pipeline and UI buttons always start in their default (OFF) state.

---

📦 Available Packages

Platform	Architecture	Editions
Windows	x64	cuda, dml (DirectML), cpu
Linux	x64 / arm64	cuda, cpu
macOS	arm64 (Apple Silicon)	mac (CPU / CoreML)

---

🔄 How to Update / Migrate

1. Download and extract the .7z archive for your platform below.
2. Launch vcclient-forge.exe (Windows) or ./vcclient-forge (Linux / macOS).
3. (Optional) To connect to a remote backend on your LAN, launch with:

   vcclient-forge --remote 192.168.x.x:18888

4. To carry over existing models and settings, copy your model_dir and settings folders into the extracted directory.

Full Changelog: v2.0.78-forge.0.1.0...v2.0.78-forge.0.1.1

v2.0.78-forge.0.1.0

VCClient Forge v2.0.78-forge.0.1.0

This milestone release introduces a native cross-platform desktop application (Windows, Linux, and macOS), Dual-UI support combining the backported Modern UI (v2.1.4) with the Classic UI (v2.0.78), a dB-scale noise gate with hysteresis, and self-contained CUDA runtime bundling on Linux.

---

🚀Highlights & What’s New

🖥️Native Desktop Application Across All Platforms (Wails v3)

• Standalone Desktop Experience: Added vcclient-forge (vcclient-forge.exe on Windows), a native desktop application powered by Wails v3.
• Clean Process Lifecycle: Closing the desktop window cleanly shuts down the backend engine—no more orphaned background processes. (Browser/headless mode remains available via start.bat / start.sh.)

🎨Dual-UI Architecture: Modern (v2.1.4) + Classic (v2.0.78)

• Modern UI Backported: Experience the sleek v2.1.4 Modern UI tailored for RVC alongside the familiar v2.0.78 Classic UI.
• Instant UI Switcher: Use the new dropdown in the top-right corner to switch seamlessly between Modern (v2.1.4) and Classic (v2.0.78) at any time.

🎙️Audio Pipeline & Recording Fixes

• dB-Scale Noise Gate with Hysteresis: Aligned the backend noise gate to the UI's decibel scale (-90 dB to 0 dB) and added a -4 dB release hysteresis to prevent voice tails from being abruptly cut off during quiet speech.
• Fixed Audio Recording & Download: Resolved server audio recording and .wav file downloads across both UIs.

📦 Portable CUDA/cuDNN Bundling on Linux

• Zero CUDA Toolkit Setup: Linux x64 and ARM64 CUDA packages now bundle portable CUDA and cuDNN shared libraries directly. Just extract and run with a compatible NVIDIA driver!

🛠️ Under-the-Hood Improvements

• Unified backend logging to eliminate string allocations on real-time audio hot paths.
• Streamlined CLI argument parsing with standard argparse and reliable two-stage Ctrl+C shutdown.
• Expanded type safety (mypy) and automated test coverage across all platforms.

---

📦 Available Packages

Platform	Architecture	Editions
Windows	x64	cuda, dml (DirectML), cpu
Linux	x64 / arm64	cuda, cpu
macOS	arm64 (Apple Silicon)	mac (CPU / CoreML)

🔄 How to Update / Migrate

1. Download and extract the .7z archive for your platform below.
2. Launch vcclient-forge.exe (Windows) or ./vcclient-forge (Linux / macOS).
3. (Optional) To carry over your existing models and settings from an earlier version or upstream v2.0.78, copy your model_dir and settings folders into the extracted directory.

Full Changelog: v2.0.78-forge.0.0.2...v2.0.78-forge.0.1.0

v2.0.78-forge.0.0.2

VCClient Forge v2.0.78-forge.0.0.2

This update fixes an issue where the FCPE pitch estimator wasn't properly downloaded or applied, and cleans up the UI for a smoother experience.

---

What’s New & Fixed

FCPE Pitch Estimator Now Fully Works!

• Fixed Model Download: Fixed an issue where the FCPE model was not being downloaded on startup. It is now automatically downloaded and verified when you start the app.
• True FCPE Audio: Fixed a bug where selecting FCPE in the settings would continue using RMVPE in the background. You can now experience genuine FCPE pitch estimation!

Cleaner Pitch Detector Menu

• The detector dropdown in the UI has been simplified to only show rmvpe and fcpe—the two best and fully supported pitch estimators. No more confusing or broken legacy options!

Stability & Quality Improvements

• Strengthened backend codebase and internal verification.

---

How to Update

Simply download the package for your platform below, unpack it, and overwrite old files!

Full Changelog: v2.0.78-forge.0.0.1...v2.0.78-forge.0.0.2

v2.0.78-forge.0.0.1

VCClient Forge v2.0.78-forge.0.0.1

Highlights

• Added native Linux ARM64 distributions for both CPU and CUDA editions, expanding support for ARM64 Linux systems such as NVIDIA DGX Spark.
• Improved backend runtime validation and enabled Python bytecode optimization level 1 in packaged builds.
• Fixed bulk performance metrics to average only over chunks that include measurements.

Available Packages

• Linux ARM64: CPU and CUDA
• Linux x64: CPU and CUDA
• Windows x64: CPU, CUDA, and DirectML
• macOS ARM64: CPU/CoreML edition, carried forward from the previous release

Windows ARM64 packages are not included in this release.

v2.0.78-forge.0.0.0

VCClient Forge v2.0.78-forge.0.0.0

The first Forge prerelease, based on w-okada/voice-changer v2.0.78, modernizes the runtime and improves voice conversion stability.

Runtime Updates

• Upgraded to Python 3.14.
• Updated NumPy to 2.5.4.
• Updated ONNX Runtime: CPU 1.31.0, CUDA 1.24.1, and DirectML 1.24.4.

Voice Conversion

• Focused the application exclusively on RVC and removed Beatrice support.
• Migrated inference to ONNX Runtime and NumPy.
• Standardized pitch estimation on RMVPE by default, with FCPE as an alternative.
• Added automatic migration of existing .pth models to ONNX at startup.

Stability Improvements

• Made queue operations in audio callbacks non-blocking.
• Fixed memory leaks and duplicate audio data in recording buffers.
• Fixed issues with disabled crossfading, division by zero, and NaN values in volume processing.
• Improved model slot updates, pitch setting synchronization, and model deletion on Windows.
• Fixed shutdown hangs and improved download hash verification.

Platform Support

Windows is the primary target. Linux and macOS support is provided on a best-effort basis. Removed ngrok and Colab integration, along with the legacy terminal UI.