Real-time voice transcription for Garry’s Mod servers
_{Opus · whisper.cpp · extensible submodule API}

Captures player voice chat via an in-process SV_BroadcastVoiceData detour, decodes Opus audio, and transcribes it in real time using whisper.cpp. Auris is a silent platform — it does nothing by default except make transcriptions available. What happens with them is up to submodule addons.

Watch Auris transcribe Garry's Mod voice chat in real time.

• Git
• Vulkan SDK 1.4.341.1 or newer (GPU builds only)
• Visual Studio with C++ build tools
• whisper.cpp model file (see "Models" below)

Download a ggml-*.bin Whisper model from:

• https://huggingface.co/ggerganov/whisper.cpp/tree/main

Place the model on your server in the Garry's Mod data folder:

• garrysmod/data/auris/ggml-tiny.en.bin

Upgrading from a previous release? Move your model file from data/whisper/ to data/auris/.

If you want to use a different model name or path, update it in:

• garrysmod_addon/auris/lua/auris/config.lua

See BUILD.md for the full build guide — all platforms, architectures, CPU vs GPU variants, and flag reference.

Quick start (Windows, CPU-only):

RUN_ONCE.bat
premake5.exe --os=windows --gmcommon=./garrysmod_common vs2022

Open projects/windows/vs2022/ and build in Release.

• Windows 11
• Vulkan SDK 1.4.341.1
• Visual Studio 18.5.11626.173

Edit garrysmod_addon/auris/lua/auris/config.lua to change the model path, language, thread count, and other options. Every key can also be overridden at runtime via ConVar (auris_threads, auris_language, auris_debug) in server.cfg — ConVar values take precedence over the file.

Auris has two transcription backends. Both fire the same Auris_Transcription hook with identical arguments, so submodules work unchanged on either.

To switch to the remote backend, fill the two keys in config.lua:

openai_api_key = "sk-...",
openai_model   = "whisper-1", -- or gpt-4o-mini-transcribe / gpt-4o-transcribe

Whisper.cpp is not loaded and the worker thread is not spawned in remote mode — the Steam voice detour still runs, captured Opus is still decoded to float32 PCM, but transcription happens entirely at https://api.openai.com/v1/audio/transcriptions. See OpenAI's speech-to-text docs for pricing and model options.

Auris exposes a hook that any addon can listen to:

hook.Run("Auris_Transcription", ply, steamid64, text, audio)

Create a new GMod addon with one server-side file:

-- lua/autorun/server/sv_myaddon_init.lua

-- timer.Simple(0) defers until all autorun files have loaded,
-- avoiding load-order issues with Auris.
timer.Simple(0, function()
    if not Auris then
        ErrorNoHalt("[myAddon] Auris core not found\n")
        return
    end

    Auris.Subscribe("MyAddon_Feature", function(ply, steamid64, text, audio)
        local name = IsValid(ply) and ply:Nick() or "Disconnected"
        Msg("[myAddon] " .. name .. ": " .. text .. "\n")

        -- optional: save the voice clip to disk
        if audio then
            file.Write("myaddon/" .. steamid64 .. ".wav", Auris.PCMToWAV(audio))
        end
    end)
end)

See API.md for the complete API, subscriber name conventions, version guards, and the publishing checklist to get your submodule listed here.

To add yours: follow the publishing guide in API.md, then open a Submit Submodule issue.

See the interactive chart on Star History (repo: ds-kimi/Auris).