Getting started

Download and Install

Get the Windows or Linux ZIP from the download section. The desktop app is free. Extract it to a writable local folder with enough space for the backend and your selected models.

Windows

1. Extract the Windows ZIP. Do not start the app from inside the archive.
2. Run Whispering Tiger.exe and accept the backend platform download.
3. Create a profile, select your audio devices, then choose models and compute devices.
4. Start the profile. The selected models download on first use.

Follow the release's CUDA requirements for NVIDIA models. audio.cpp with Vulkan uses the graphics driver and does not require CUDA.

Linux

The Linux download targets x86-64 with glibc 2.36+. The UI needs an OpenGL-capable X11/XWayland desktop and PulseAudio or PipeWire's PulseAudio compatibility service.

Extract the Linux ZIP, open a terminal in that folder, and run:

chmod +x whispering-tiger-linux-amd64
./whispering-tiger-linux-amd64

Run as your normal desktop user. Accept the Linux backend download, then create a profile. The packaged CUDA backend includes its CUDA runtime libraries; GPU use still requires a compatible NVIDIA driver. audio.cpp offers CPU/Vulkan on Linux.

Which Hardware and Runtime Should I Use?

Select a device separately for STT, translation, TTS and OCR. Support for one runtime does not imply support for all models.

AMD/Intel: choose audio.cpp, a supported STT/TTS model, Vulkan and your GPU. Translation and OCR can use a different device, such as CPU. Vulkan needs a compatible driver, not the Vulkan SDK.

Memory: start with a small model. STT, translation and TTS add their memory requirements when loaded together. Lower precision can reduce weight memory; speed and supported precisions depend on the runtime and device.

HIP/ROCm requires a custom audio.cpp server build and is not a managed UI option. DirectML is not an audio.cpp backend.

Microphone, Game Audio and TTS

For microphone transcription, select your input and test its level in the profile. For desktop audio, use WASAPI loopback on Windows or a PulseAudio/PipeWire monitor source on Linux.

To send TTS into VRChat or another voice-chat application, route it through a virtual audio device or Linux audio tools. Keep generated speech out of the input being transcribed to avoid a feedback loop.

• Windows: select the virtual cable as Whispering Tiger's audio output and as the voice-chat app's microphone.
• Linux: send Whispering Tiger's output to a virtual sink and select its monitor as the voice-chat input, or connect the streams with PipeWire routing tools.
• Hear TTS yourself: enable tts_use_secondary_playback in Advanced > Settings and choose your headphones as the secondary playback device.

Offline Use

Download the backend, runtimes and selected models first. Local inference can then run offline. Online plugins need a connection and may send data to their provider. Some model options require manually supplied local weights; check the model instructions.

Set Up Text-to-Speech

1. Choose the TTS type and compute device in your profile.
2. Open Text-to-Speech, select a model and voice, then generate a short test sentence.
3. If your model supports voice cloning, use Open Voice Directory in the voice selector to add a reference recording.

VRChat and Streaming Overlays

For VRChat chatbox text, enable OSC in VRChat and use its address as the OSC destination in Whispering Tiger. On the same PC, that address is 127.0.0.1. This sends text; spoken TTS uses the audio routing described above.

For streaming overlays, enable the WebSocket server and add an overlay HTML file from the backend's websocket_clients folder as a browser source in your streaming application. Keep the default local WebSocket address when both applications run on the same PC.

Realtime Transcription

Enable realtime mode for intermediate results while you speak. Start with a small model and a supported GPU. If updates lag, use a smaller model or increase realtime_frequency_time in Advanced > Settings to request updates less often.

Whisper-specific controls do not apply to every model. Use the profile's memory estimate to check whether your combined STT, translation and TTS models fit.

Install Plugins

Use the app's plugin installer to browse and install extensions. Configure the installed plugin in the Plugins tab; its settings depend on the plugin.

For manual installation, place the plugin's .py file in the Plugins folder and restart the app.

Need More Help?

Check the log under Advanced if a model fails to load or an audio device does not work. For help with your setup, visit community support.