v0.5.3

[0.5.3] - 2026-10-04

Changed

• Default source builds of the standalone CLI, GUI, and shared runtime now use
  Windows ML without enabling native TensorRT. TensorRT remains an explicit
  build feature; RNNoise and GTCRN remain enabled by default.

Fixed

• Stabilized SOLA and PSOLA sliding-window energy calculations to avoid
  rounding errors affecting overlap alignment.
• Normalized integer WAV input using its original PCM bit depth.
• Preserved loaded VST3 models across host resets, processing-mode changes, and
  block-size changes. Offline export/freeze preserves the initial input and final
  partial chunk through latency compensation and tail processing.
• Corrected native TensorRT buffer initialization, engine-build serialization,
  and compatibility with supported RVC export conventions.
• Corrected streaming denoiser priming delay and RVC output candidate lengths,
  including the matching fixed GPU input profiles.
• Recovered from transient exclusive WASAPI capture errors and stopped sessions
  on fatal audio-stream errors while preserving diagnostics.
• Preserved still-voiced output tails after silent input chunks and compensated
  independent capture/render device clock drift with continuous output resampling.
• Delay-loaded native TensorRT DLLs in the GUI so an explicitly enabled combined
  build can start before those DLLs are needed.
• Included packaged .vst3 module binaries in the release path-leak scan and
  retained explicit TensorRT coverage in the release Clippy gate.

Distribution notes

• Prefer release builds for realtime use. The device-clock resampler has
  substantially higher CPU cost in unoptimized debug builds.
• Windows binaries are not code-signed; Windows may display a security warning
  when downloading or running them.

v0.5.2

[0.5.2] - 2026-09-19

Added

• Explicit OpenVINO CPU, GPU, and NPU selection in the CLI, GUI, and VST3,
  with device discovery and guided first-time setup in the graphical interfaces.
• Optional CLI inference performance reports and WAV benchmarking tooling.

Changed

• Improved VST3 control layout and displayed decibel parameters to two decimal
  places, preserving parameter IDs and saved settings.
• Made English the default README and expanded English/Japanese backend and
  hardware support documentation.
• Optimized OpenVINO GPU ContentVec loading with fixed input dimensions while
  keeping RVC input shapes dynamic.

Fixed

• Corrected Windows ML automatic session fallback.
• Routed RMVPE to OpenVINO CPU when using OpenVINO GPU or the unrestricted
  OpenVINO provider to avoid incorrect pitch inference on the GPU.
• Corrected local VST3 bundle naming for side-by-side backend variants.

Distribution notes

• OpenVINO NPU selection is available but has not been validated on NPU hardware.
  GPU selection does not guarantee that every operation runs on the GPU.
• Windows binaries are not code-signed; Windows may display a security warning
  when downloading or running them.

v0.5.1

[0.5.1] - 2026-09-13

Added

• Guided GUI setup for models, execution providers, and audio device testing.
• Recently selected audio devices are prioritized in device pickers.
• GUI interaction tests for both isolated distribution backend variants.

Changed

• Redesigned the main GUI controls, simplified status display, and moved setup
  and language controls into scrollable settings.
• Added clearer error highlighting and colored audio level meters.

Fixed

• Preserved model export metadata during built-in PTH-to-ONNX conversion.
• Cleared stale GUI errors after a successful stop.

Distribution notes

• Windows binaries are not code-signed; Windows may display a security warning
  when downloading or running them.

v0.5.0

[0.5.0] - 2026-09-07

Added

• Built-in RVC PTH-to-ONNX conversion with GUI integration.
• Support for rvc-onnx-web streaming exports, including NSF phase and noise
  inputs, on native TensorRT and Windows ML TensorRT RTX.
• Realtime worker stop causes in engine status and usable execution providers
  from the live Windows ML catalog in provider pickers.

Changed

• Windows ML packages now require Windows App SDK Runtime 2.1 or newer and use
  ONNX Runtime API 24 through ort 2.0.0-rc.13.
• Upgraded native TensorRT to 11.2.1 with CUDA 13.3 Update 1 and isolated native
  caches by SDK version. Existing engines rebuild on first use after upgrading.
• Consolidated provider abstraction in the shared core and updated dependencies.
• Release publishing now requires the CI formatting and Clippy checks to pass.

Fixed

• Resolved the Windows ML bootstrapper beside the VST3 plugin instead of
  depending on the DAW's DLL search policy.
• Corrected resampling timelines and preserved final WAV tails.
• Corrected NSF phase carry across overlapping windows and kept RVC latent
  noise on an absolute frame timeline.
• Bounded PTH ZIP parsing with the zip crate and updated dependencies for
  security fixes.
• Deferred audio stream error logging away from the realtime callback.

Performance

• Reduced fixed-hop resampler latency and reused RMS scratch buffers.

Distribution notes

• Windows binaries are not code-signed; Windows may display a security warning
  when downloading or running them.

v0.4.0

[0.4.0] - 2026-06-25

Added

• Standalone GTCRN input denoising across the shared runtime, CLI, GUI,
  Windows ML packages, and native TensorRT path.
• Optional ASIO support for standalone CLI and GUI builds, with per-direction
  audio host selection and RT-priority CPAL handling.
• Microsoft Store GUI-only MSIX packaging via the Windows App Development CLI.
• Offline chunk-join sweep tooling, model-free pipeline benchmarks, and
  validation shared across frontends.

Changed

• Reworked the standalone audio host layer to support separate input and output
  backends.
• Optimized mono channel mixing, RMS mix-gain processing, output level
  postprocessing, and shared waveform handling.
• Expanded distribution packaging and release checks for GTCRN, Store MSIX, and
  version-specific GitHub release notes.

Fixed

• Supported Applio-style RVC exports and models with the optional rnd
  latent-noise input.
• Accepted f0: true metadata and sized model windows at the exported model's
  sample rate.
• Failed fast when the Windows ML bootstrap DLL is missing and handled catalog
  execution-provider preparation consistently.
• Fixed TensorRT-only RVC profile builds.
• Validated conversion settings consistently across frontends.

v0.3.0

[0.3.0] - 2026-06-15

Added

• Live passthrough switching for standalone sessions with a complete model set.
• Named GPU device selection for supported CUDA and TensorRT providers.
• Staged model-loading progress in the GUI.
• Chunk-join diagnostics for measuring output-boundary artifacts.
• Shared conversion-pipeline architecture guidance, CI checks, cargo-deny
  policy, a check-only pre-commit rustfmt hook, and CPU hot-path benchmarks.

Changed

• GPU Priority now applies to every backend, not just native TensorRT: it sets a
  process-wide Windows GPU scheduling priority class (in addition to the
  TensorRT CUDA stream priority on that path), so the control is shown in the GUI
  for Windows ML / CPU builds as well. High additionally opts the process out of
  CPU power throttling (EcoQoS) so inference keeps full clock when the window is
  in the background, removing the large foreground/background timing difference.
• CLI, GUI, VST3, and WAV conversion now reuse shared chunk-conversion,
  smoothing, and output-assembly components.
• Standalone realtime processing now wakes the input worker when audio arrives
  instead of polling every 2 ms.
• The GUI chunk-size control now supports values down to 40 ms.
• Updated CPAL, nice-plug, egui, rfd, toml, rubato, and compatible transitive
  dependencies.

Fixed

• Reduced audible chunk-join artifacts at small chunk sizes.
• Restored VST3 processing correctly after plugin reload.
• Opened CPAL streams using the device's native channel count.
• Surfaced Windows ML catalog execution-provider preparation failures.
• Ensured VST3 installation stages the required runtime DLLs.

Performance

• Removed repeated allocation and redundant work from inference, DSP, and
  SOLA/PSOLA hot paths.
• Vectorized SOLA/PSOLA offset search and reused input-side inference buffers.

v0.2.1

Added

• Standalone RNNoise input denoising in the GUI and CLI.
• Configurable input noise gate before RVC feature and F0 extraction, available
  in the standalone apps and VST3 plugin.
• Optional F0 post-processing support in the core RVC pipeline.
• On-demand download of explicitly selected Windows ML catalog execution
  providers during standalone GUI and CLI model loading.
• Deterministic CPU-only A/B audio comparison tooling for regression analysis.

Changed

• Release publishing now relies on GitHub's asset digests instead of generating
  separate SHA-256 sidecar files.

Windows Warning

The distributed Windows binaries are not code-signed. Windows may display a
SmartScreen or unknown-publisher warning when launching them.

v0.2.0

What's new in 0.2.0

Added

• Standalone GUI app (vc-gui.exe) backed by a shared realtime runtime, shipped
  alongside the CLI in the standalone packages.
• doctor CLI command for runtime diagnostics.
• TensorRT GPU priority control.

Changed

• Standalone packages now bundle the GUI together with the CLI.
• Refined GUI runtime controls and diagnostics.
• Capped the TensorRT builder at 4 max threads.
• Distribution packaging now generates exact per-binary Rust license notices.
• TensorRT packages always bundle every GPU builder resource for full
  compatibility.

Fixed

• Preserved silent output buffering in the realtime worker.

Downloads

Package	Backend	For
vc-rs-windowsml	Windows ML	GUI + CLI, broad GPU support (needs Windows App SDK Runtime)
vc-rs-tensorrt	TensorRT	GUI + CLI, NVIDIA GPU
vc-vst3-windowsml	Windows ML	VST3 plugin, broad GPU support
vc-vst3-tensorrt	TensorRT	VST3 plugin, NVIDIA GPU

GitHub shows a SHA-256 digest for each asset above (click the asset's details);
verify a download with Get-FileHash -Algorithm SHA256 <zip> and compare.

Note

These binaries are not code-signed, so Windows SmartScreen / Defender may show
a warning on first launch. Choose More info → Run anyway to proceed.

vc-rs v0.1.0

First release of vc-rs, a Rust RVC (Retrieval-based Voice Conversion) voice changer for Windows (x64). Real-time mic→speaker and WAV→WAV via the CLI, plus a VST3 plugin for DAWs.

Downloads

Two inference backends, CLI and VST3 each:

Package	Backend	Notes
vc-rs-cli-windowsml	Windows ML	Smallest. Needs Windows App SDK Runtime 2.x (provides ONNX Runtime + DirectML). Broad GPU support.
vc-vst3-windowsml	Windows ML	VST3 plugin, same runtime requirement.
vc-rs-cli-tensorrt	TensorRT	NVIDIA-only, self-contained (runtime DLLs bundled, all GPU architectures sm75–sm120). Needs an up-to-date NVIDIA driver.
vc-vst3-tensorrt	TensorRT	VST3 plugin, self-contained NVIDIA build.

Models

Models are not bundled. Each package includes download-models.ps1 to fetch the reference ContentVec + RMVPE models into .\assets\. Supply your own RVC voice model (.onnx). See the included INSTALL.txt.

Install

• CLI: unzip, run .\vc-rs.exe --help (keep the DLLs beside the exe).
• VST3: copy the .vst3 bundle into a VST3 search path (e.g. %CommonProgramFiles%\VST3).

Third-party license notices are included in each package. MIT licensed.