excavate-listen

The speech-to-text runner for Excavate, a local search app for Windows. Excavate downloads it, then uses it to make recordings searchable by what's said in them.

Why it exists

• whisper.cpp publishes no Windows build for Vulkan, and its Windows CUDA build leaves out cuBLAS, so it runs on the processor.
• Excavate needs one small program that runs Whisper and Parakeet on any graphics card or on the processor, keeps a model loaded, and takes sound over a pipe.

These files are that program, built from whisper.cpp's source. They contain no model weights. The models come from ggml-org's own repositories: ggerganov/whisper.cpp, ggml-org/parakeet-GGUF and ggml-org/whisper-vad.

Files

File	What's in it
excavate-listen-b5454-3-vulkan.zip	Build 3: excavate-listen.exe (Silero's voice detector first, for Whisper and Parakeet), whisper.cpp's libraries, every processor variant, Vulkan, Microsoft's C++ runtime (18 MB). Every program and library carries version details (company, product, description, version 1.5454.3.0, license)

Excavate checks each file's SHA-256 before using it.

Built from

• whisper.cpp tag b5454 (commit d1be6fd), with its backends as loadable libraries (GGML_BACKEND_DL).
• The listener's source and its build script are tools/listen/listen.cpp and tools/listen/build.bat in Excavate's repository.
• CUDA's library is built for Turing to Blackwell (GTX 16 and RTX 20 onwards).

Licenses

• whisper.cpp and ggml: MIT (LICENSE-whisper.cpp.txt).
• The listener: MIT (LICENSE-excavate-listen.txt).
• NVIDIA's runtime and cuBLAS are redistributed under NVIDIA's CUDA Toolkit EULA (EULA-NVIDIA-CUDA.txt).
• Microsoft's C++ runtime is redistributed under Visual Studio's terms for app-local use.