Use whisper.cpp through StarWhisper's Windows interface. Record, stop, and insert text into supported apps. Download a model for local processing, with CPU, compatible CUDA, or optional Vulkan support. No Python setup for the app.
Free: Tiny, Base, and Small local models, 500 words/day and 3,500 words/week. Pro: $10/month or $80/year for unlimited transcription, larger models, and file transcription.
For the open-source engine, command-line tools, or build instructions, start at the official whisper.cpp repository. Follow its Windows instructions and choose the model and backend required by that build. StarWhisper is a separate desktop application that uses whisper.cpp; this is not the upstream project's download site.
For a Windows dictation interface, try StarWhisper. It provides a floating widget, a configurable recording hotkey, model downloads in Settings, and insertion into supported text fields. Record your sentence, stop, then wait for transcription. Start with the Free local models before deciding whether you need Pro for larger models or saved-file transcription. The Whisper desktop app guide also explains the difference from Const-me's separate WhisperDesktop application.
whisper.cpp implements OpenAI's Whisper speech recognition model in C/C++. It can run transcription without the Python Whisper runtime. Its upstream command-line tools still need a compatible build, a downloaded model, and the options appropriate to your audio and hardware.
StarWhisper packages that local engine with its own Windows interface and setup controls. The GUI workflow removes the need to type transcription commands; it does not remove the need for a working microphone, enough model memory, or compatible GPU drivers. CPU inference remains available when a GPU backend cannot be used.
This page covers what whisper.cpp does on Windows, how StarWhisper packages it into a desktop application, what hardware configurations are supported, and how whisper.cpp compares to running Whisper in Python on Windows.
StarWhisper's Free tier includes Tiny, Base, and Small local models. Pro adds Medium, Large-v3-Turbo, and Large. The model files use the whisper.cpp format and are managed through the Settings panel, with approximate downloads from 75MB for Tiny to 2.9GB for Large.
whisper.cpp can use NVIDIA CUDA when a compatible GPU and driver path are available. StarWhisper also has an optional Vulkan path for supported NVIDIA, AMD, and Intel hardware when its probe succeeds. If no GPU backend is available, the app falls back to CPU inference. Actual speed depends on the selected model, hardware, drivers, audio, and settings.
The same model file handles many Whisper languages. Language specification is a runtime parameter, not a separate model download. whisper.cpp also implements Whisper's translation mode, which transcribes audio in one language while outputting English text. StarWhisper exposes language selection and translation options where supported by the app.
The whisper.cpp engine supports timestamp output. When a StarWhisper Pro file transcription produces subtitles, its saved result offers SRT and VTT downloads. Review subtitle wording and timing, which depend on the model, audio, and recording conditions.
CPU-only inference is supported, with performance depending on the processor and selected model. On compatible Windows hardware, the optional Vulkan path can provide another backend; otherwise the app uses CPU fallback.
StarWhisper ships a pre-compiled whisper.cpp binary for Windows x64 and packages the Windows runtime needed for its GUI workflow. Installation does not require Python, Conda, pip, or any package manager. Setup time depends on the installer, model downloads, and the Windows machine.
StarWhisper's user interface is built with Electron, which provides the Windows GUI layer. The Electron frontend communicates with the whisper.cpp backend process through a local IPC mechanism. From the user's perspective, this presents as a standard Windows application, while the backend process runs separately from the GUI.
CUDA-enabled builds of whisper.cpp require CUDA runtime libraries. StarWhisper packages the runtime path used by its Windows build, so users do not need the full CUDA toolkit for the app workflow. A compatible NVIDIA driver still needs to be present, and the exact result depends on the installed driver and package.
Whisper model files are managed through the Settings panel. The current app package lists approximate files of 75MB (Tiny), 142MB (Base), 466MB (Small), 1.45GB (Medium), 1.55GB (Large-v3-Turbo), and 2.9GB (Large). Users can choose the models they need instead of downloading every file.
StarWhisper keeps the whisper.cpp worker process running with the selected model loaded, so individual transcription requests do not need to reload the model each time. The resulting response time still depends on model size, hardware, drivers, audio, and settings.
Use these model-selection cues when choosing a whisper.cpp model on Windows. They are guidance, not performance promises.
| Model | CPU-only guidance | GPU guidance | Download size |
|---|---|---|---|
| tiny | Fastest CPU choice | GPU optional | ~75MB |
| base | Balanced CPU choice | GPU optional | ~142MB |
| small | Balanced choice; CPU may be slower | GPU can help | ~466MB |
| medium | May be slow on CPU | Use a supported GPU backend | ~1.45GB |
| large-v3-turbo | Usually not practical on CPU | Needs more memory; check Settings | ~1.55GB |
| large | Usually not practical on CPU | Needs more memory; check Settings | ~2.9GB |
Actual fit and speed vary with CPU/GPU generation, available memory, drivers, audio characteristics, model settings, and system load. Check the Settings guidance before choosing a larger model. The whisper.cpp GitHub repository has community benchmarks for a wider range of hardware configurations.
Start with Tiny, Base, or Small. These are the Free local models and make practical starting points for live dictation. Check the Settings guidance before choosing Medium, Large-v3-Turbo, or Large, because larger models need more memory and may be slower on CPU-only hardware.
Use Medium, Large-v3-Turbo, or Large when you need a larger model and your hardware can handle the extra memory and download. For short dictation, compare results with a smaller model and use the Settings guidance if you are unsure.
When a compatible NVIDIA driver and CUDA path are available, StarWhisper can use CUDA. The optional Vulkan path may also work on supported NVIDIA, AMD, and Intel hardware when its probe succeeds. If no GPU backend is available, the app falls back to CPU. Check Settings before choosing Medium, Large-v3-Turbo, or Large, since memory needs and speed vary by model and hardware.
StarWhisper targets Windows 10/11 x64 and supports CPU-only use. Compatible NVIDIA hardware can use CUDA, and supported Vulkan hardware may use that backend when the probe succeeds. Larger models need more memory, so use the Settings guidance and start with a smaller model when hardware is limited.
The whisper.cpp engine is written in C/C++ and does not need the Python Whisper runtime for transcription. Using the upstream tools still requires their documented setup and a model download. StarWhisper packages its own Windows GUI workflow, so you can install the app and download a model through Settings without Python or command-line setup.
StarWhisper can use NVIDIA CUDA when a compatible driver path is available. It also has an optional Vulkan path for supported NVIDIA, AMD, and Intel hardware when its probe succeeds. If no GPU backend is available, the app uses CPU inference.
whisper.cpp uses the same Whisper model families, but output can vary with the selected model, backend, hardware, audio, and settings. Compare the available models in Settings for the workflow you need.
Approximate model files are Tiny 75MB, Base 142MB, Small 466MB, Medium 1.45GB, Large-v3-Turbo 1.55GB, and Large 2.9GB. Additional models are downloaded on demand from the Settings panel.
StarWhisper makes whisper.cpp accessible on Windows without Python, command-line work, or CUDA toolkit installation. The Free plan needs no account, and Pro is $10/month or $80/year for unlimited use and larger models.