Local AI for Speech, Text and Screen Content
Whispering Tiger runs on Windows and Linux. Use it for microphone transcription, game translation, spoken output, VRChat chatbox text or streaming overlays. Choose the models and devices separately for each task.
Languages Depend on the Model
Whisper supports multilingual transcription and speech translation into English. For other target languages, follow STT with a text translator: NLLB-200 supports 200 languages, M2M-100 supports 100, Hunyuan MT supports 33, and MiLMMT-46 supports 46. TTS and OCR have separate language lists.
Speech-to-Text (STT / ASR)
- Whisper, Faster Whisper/CTranslate2 and Transformers Whisper, including custom models.
- Qwen3-ASR, Seamless M4T, MMS, SpeechT5, Wav2Vec-BERT, NeMo Canary/Parakeet, VibeVoice-ASR and Voxtral.
- audio.cpp GGUF models, including Qwen3-ASR, Nemotron, VibeVoice, Voxtral Realtime and Kroko.
- Microphone/desktop capture, file transcription, voice activity detection, push-to-talk and intermediate realtime results.
- Language detection and timestamps where supported by the model. Word timing is not available from every backend.
Text and Speech Translation
Translate typed text or transcription results with NLLB-200, M2M-100, Seamless M4T, Hunyuan MT, MiLMMT-46 or Voxtral. Combine speech recognition, translation and TTS for spoken output in another language. Optional plugins connect online services such as DeepL.
Text-to-Speech and Voice Cloning
Integrated engines include Silero, Chatterbox, F5/E2, Kokoro, Zonos, IndexTTS 2.5, Qwen3-TTS, Orpheus and Maya1. Preset voices, cloning, emotion and streamed playback depend on the model.
audio.cpp adds native GGUF models such as Supertonic 3, MagpieTTS, DotTTS, IndexTTS, OmniVoice and VoxCPM. Use Vulkan for supported AMD/Intel GPUs, CUDA on Windows for NVIDIA, or CPU.
Route speech to your chosen playback device and optionally a second device. TTS voice-conversion plugins can process generated speech; whole-utterance conversion delays playback until generation and conversion finish.
OCR and Screen Text Translation
Extract text from images or screen captures using EasyOCR or GOT-OCR 2.0, then translate it. Tesseract is available through a plugin. Screen and window capture behavior depends on the operating system and desktop session.
Desktop Audio and Routing
- Windows: WASAPI loopback for desktop audio and process loopback for individual applications.
- Linux: PulseAudio or PipeWire's PulseAudio compatibility service, including monitor sources for desktop audio.
- Additional sources: process more than one audio input and configure each source's translation and outputs.
- TTS into voice chat: use a virtual audio device on Windows or route streams through Linux audio tools.
VRChat, Overlays and Profiles
Send text to VRChat's chatbox via OSC or to browser clients over WebSockets. Browser overlays can display results in streaming applications. Save separate profiles for your microphone, games, translation languages and hardware settings.
Plugins
Install plugins through the UI for subtitles, subtitle export, voice conversion, online services and other integrations. Available features and requirements depend on the plugin.
CPU, NVIDIA, AMD and Intel Support
PyTorch/Transformers and CTranslate2 models use CPU or supported NVIDIA CUDA devices. audio.cpp supports Vulkan for AMD, Intel and NVIDIA STT/TTS. Its managed Linux runtime provides CPU/Vulkan; the Windows runtime also offers CUDA. Vulkan support does not extend to every model or task.
Model size, precision, available memory and concurrent tasks affect latency. Start with a smaller model and use the profile's memory estimate as a guide.
Local and Offline Processing
Local models process audio, text and images on your machine. Download the backend, runtimes and selected weights before using them offline. Updates, new model downloads and online plugins need a connection. Online plugins may send data to their provider.
The UI and backend are open source. Model weights and plugins have separate licenses; some model options require manually supplied weights.