Callout

Translates in-game voice chat from any language into English, live, for any game: Valorant, CS2, Apex, Fortnite, or anything else that plays voice through Windows. Lines appear in a chat window or an overlay on top of the game. Everything runs locally on your GPU. No account, no API key, no cost.

Download

These links always point to the newest release:

	Installer (recommended)	Portable
Windows x64 (most gaming PCs)	Callout-Setup-x64.exe	Callout-win-x64.zip
Windows on ARM	Callout-Setup-arm64.exe	Callout-win-arm64.zip

• Installer: installs for your user only (no administrator prompt), adds a Start menu entry, and can be removed from Settings → Apps. Uninstalling also deletes the downloaded models.
• Portable: unzip anywhere and run Callout.exe.

All versions are on the Releases page. Builds are not code-signed yet, so Windows SmartScreen may warn on first launch: choose More info → Run anyway.

Callout is in early access (0.x). Expect rough edges, and please report what you find.

How it works

Game audio (one app, a whole device, or VB-Cable for voice only) ──► Callout ──► your headset (pass-through for VB-Cable)
                                         │
                                         ├─ splits speech into lines (loudness, then Silero VAD)
                                         ├─ Whisper detects the language and translates it → English text (GPU via Vulkan)
                                         └─ shows each line in the chat window or the in-game overlay

Languages: auto-detect (the default) works out the language of each line separately, across the ~99 languages Whisper knows. A Russian, a Turk and a Pole in one lobby each get translated, and every line is tagged (RU, TR, PL). Lines already in English are hidden by default. Pin a language under Teammates speak if short callouts get misdetected.

Display modes (switch any time, even while listening):

• Chat window (default): a normal window with scrollback and timestamps. Put it on a second monitor, or pin it on top.
• Game overlay: the last few lines float over the game and fade after 12 s. Clicks pass through when locked.

Both are ordinary windows. Callout never reads, injects into or hooks the game, so it doesn't interact with anti-cheat.

Setup

Pick what to listen to under Voice chat source:

Source	Install?	Hears	Notes
App: your game	Nothing	Only that game	Start the game first, then press refresh (⟳) so it appears. Discord, music and browsers are left out. Windows 10 2004+.
App: Discord (or another voice app)	Nothing	Only that app	Use this if your team talks over Discord rather than in-game chat.
Everything playing on speakers	Nothing	All PC audio	The default. Simplest, but music and videos get translated too.
CABLE Output	VB-Cable	Voice chat only	Best accuracy; see below.

None of the options except VB-Cable ask for microphone access.

Best accuracy (optional, about 2 minutes): a game plays voice chat and gunfire from the same process, so per-app capture hears both. Most games let you send voice chat to a separate output device, which gives the translator voice only:

1. Install VB-Cable (free): https://vb-audio.com/Cable/ and reboot.
2. In your game's audio settings, set the voice chat output device to CABLE Input. In Valorant: Settings → Audio → Voice Chat → Output Device.
3. In Callout, the source switches to CABLE Output automatically. Keep Hear voice chat on set to your headset so you still hear your team.

Windows lists VB-Cable as a microphone, so it asks for microphone access in this mode. Your real mic is never used.

Overlay: run the game in windowed or borderless fullscreen (exclusive fullscreen hides every overlay), choose Game overlay, untick Lock overlay, drag it into place, then lock it again.

The first start downloads the chosen model (~540 MB for Balanced) to %LocalAppData%\Callout\models.

Resource use

• Idle costs nothing: the GPU only works while someone is talking. Silence never reaches the model.
• Small per-line cost: the encoder only processes as much audio as a line can hold (~13 s), not Whisper's usual 30 s window, and flash attention is on.
• Stop frees everything: the model is loaded on Start and released on Stop, so RAM and VRAM go back to the game.
• The game comes first: the app runs at below-normal CPU priority, and audio capture keeps its real-time thread.
• Bounded memory: if translation falls behind, stale lines are dropped. Chat history is capped at 300 lines. The non-concurrent GC means no background GC thread.
• Fast start: ReadyToRun-compiled, so startup doesn't spend CPU on JIT.

Privacy

Audio never leaves your PC: capture, speech detection and translation all run locally. Callout's only network traffic is downloading the speech models from Hugging Face the first time you start a quality level. Nothing is uploaded, and there is no telemetry.

Code signing policy

Signing is being set up with SignPath Foundation; until it's active, releases are unsigned. Once active:

• Release binaries are built and signed only by this repository's GitHub Actions (release workflow) from tagged commits. No release is signed from a developer machine.
• Only binaries built from this repository's source are signed (Callout.exe, Callout.dll, Callout.Core.dll, and the installers). Bundled runtime and library files keep their original publishers' signatures.
• Committers and reviewers: @iStorry. Release approver: @iStorry.

Setup steps for maintainers are in docs/code-signing.md.

Requirements

Build	Runs on	Translation runs on
x64	Windows 10/11 on Intel/AMD	GPU via Vulkan (NVIDIA, AMD or Intel; just the driver), CPU fallback
ARM64	Windows 11 on ARM (Snapdragon)	CPU only: Whisper has no ARM GPU backend yet. Use the Fast quality setting.

Build

dotnet test

./scripts/publish.sh

This produces the portable artifacts/Callout-win-x64.zip and artifacts/Callout-win-arm64.zip. The installers are built in CI with Inno Setup from installer/Callout.iss, since that step needs Windows.

Releasing

Push a version tag. GitHub Actions tests, builds both architectures, builds both installers, and publishes the release with all four files attached:

git tag v0.2.0 && git push origin v0.2.0

Versions below 1.0 are titled early access. To check a release build without publishing, run Actions → Release → Run workflow: the installers and zips are attached to that run as artifacts. Every push to main also runs the format check, build and tests (CI).

Tuning

%LocalAppData%\Callout\settings.json → Segmenter:

• SpeechRmsThreshold: lower it if quiet teammates are missed, raise it if mic noise produces junk lines.
• SilenceToEnd: how long a pause ends a line.
• MaxUtterance: long speeches are cut into pieces of at most this length.

Layout

Path	What
src/Callout.Core	Platform-free logic: speech segmentation, transcript cleanup, audio sources, settings
src/Callout	WPF app: audio capture (NAudio), Whisper.net translation, chat window and overlay
tests/Callout.Core.Tests	Unit tests for the core
installer/Callout.iss	Inno Setup installer script
scripts/publish.sh	Builds the portable zips

License

MIT. Third-party components and their licenses are listed in THIRD-PARTY-NOTICES.md.