Real-time local speech-to-text for Windows. Two language panes side by side, one model each, running entirely on your machine. No cloud, no telemetry, no account, no API keys, no network at runtime.

Built for meetings, interviews, lectures and interpretation — the situations where a transcript needs to keep up with people talking, and where the audio should never leave the machine. Captures what the machine is playing via WASAPI loopback, so it works on calls, videos and streams without a virtual cable.

Accuracy is the design goal. Where a trade-off exists between staying current and keeping every word, this program keeps the word.

Note on how this was built. OLAS was heavily vibe-coded: roughly 95% of the code was written by different agentic AI models, with a human directing the design, testing on real hardware and deciding what shipped. The measurements in PATCHNOTES.md exist because the AI-written parts got things confidently wrong more than once, and only measurement caught it. Treat the code accordingly: it works, but it has not had a conventional human review.

• Two languages at once — English and Spanish, each in its own pane with independent Start/Stop, collapse and decouple-to-window controls
• Two modes — Normal (more accurate) and Potato (ultralight), chosen on first run and changeable from Options
• Focus mode — hides the toolbar and pane headers so the transcript fills the window
• Live transcript file — written line by line to olas-moonshine-notes.txt
• Options popup — capture device, English model, zoom, timestamps, light/dark theme, auto-scroll
• Update check — asks GitHub whether a newer release exists
• Diagnostics — -v writes per-line latency to olas-debug.log, --stats prints inference diagnostics on exit

Which mode runs is decided by the English model. Spanish is always Small Streaming: no Medium Spanish model exists.

The English model is chosen on first launch and remembered in olas-model.txt beside the executable. To change it later: Options → English model, then restart.

1. Open the Releases page.
2. Download the latest OLAS-win64-1.1.x.zip.
3. Extract it anywhere.
4. Double-click OLAS.bat.

The zip contains everything: the executable, the runtime DLL, and all three models. Nothing else to install.

Requirements

• Windows 10 21H2 or newer (Windows 11 recommended)
• x64 CPU with AVX2 — Intel Haswell (2013) or AMD Excavator (2015) and newer
• 4+ logical CPU threads recommended
• ~800 MB free disk

No GPU is used or required.

WASAPI loopback (miniaudio)
        |  16 kHz mono s16, 50 ms chunks
        v
capture.c ring buffer
        |  capture_read_chunk()
        v
capture thread
        |  Engine::feed() -> AudioQueue per language (lossless, non-blocking)
        v
worker thread per language
        |  Moonshine Transcriber (streaming)
        v
listener -> Win32 RichEdit, and to olas-moonshine-notes.txt

Two models run in parallel, each processing the same audio stream. Both transcribe everything; you read the pane for the language being spoken. This is simpler and more robust than language detection, at the cost of the non-target pane producing nonsense — a known limitation, discussed in PATCHNOTES.md.

Each language runs on its own worker thread with its own Transcriber, on its own CPU core set. The core budget and per-model split are derived from the detected hardware at startup.

src/
    main_win.cpp         entry point, CLI, capture thread
    moonshine_engine.*   Engine, worker slots, Streaming/NonStreaming workers
    resource_plan.*      CPU detection, per-model core split, affinity
    model_choice.*       persisted English model selection
    win32_ui.*           panes, toolbar overlay, buttons, popup menu
    capture.c / .h       WASAPI loopback + ring buffer
    miniaudio.h          vendored miniaudio 0.11
    update_check.*       GitHub release check
    transcript_sink.h    engine <-> UI abstraction
tools/
    fetch-streaming-models.ps1
cmake/
    copy-models.cmake    bundles models into the release zip

olas-win.conf holds the transcription parameters:

[general]
vad_threshold = 0.5             # speech/silence threshold
vad_max_segment_duration = 12   # longest single line, seconds
transcription_interval = 1.0    # how often the decoder re-runs

Delete it and built-in defaults apply. Bad values and unknown keys warn on stderr; they never stop the program starting.

Threading, affinity and the model choice are not configured here — they come from the hardware probe and from olas-model.txt.

3=BaseStreaming is rejected — declared in Moonshine's C API for forward-compatibility but not supported.

• Moonshine Voice — the speech-to-text models and C++ runtime. Small and Medium Streaming architectures, MIT licensed.
• Win32 / RichEdit — the user interface, with a custom button class for theme-aware controls.
• miniaudio — WASAPI loopback capture, vendored single-header.
• ONNX Runtime — inference, built as part of Moonshine's CMake project.

Prerequisites: Visual Studio 2022 Build Tools with "Desktop development with C++", CMake ≥ 3.20, Git, PowerShell 5.1+, and optionally Ninja.

git clone https://github.com/Igna-Mendez/Open-Local-Audio-Scribe-OLAS.git
cd Open-Local-Audio-Scribe-OLAS
powershell -ExecutionPolicy Bypass -File setup.ps1

mkdir build
cd build
cmake -G Ninja -DCMAKE_BUILD_TYPE=Release ..
cmake --build . --config Release -j

setup.ps1 checks the toolchain, fetches all three models into models\, and prints the exact configure command.

The first configure downloads the Moonshine SDK — a prebuilt release archive (~26 MB) containing the headers, the import libraries and onnxruntime.dll. Nothing is built from source, so configure takes a minute or two rather than the 5–15 minutes a source build of ONNX Runtime would need.

It follows Moonshine's latest release, and stays current: the configure records which release it fetched, checks for a newer one each time, and re-downloads when there is one. Reconfigure to pick up a new release.

Deleting build\moonshine-sdk\ also forces a fresh download. Offline, or if the release check fails, the cached SDK is kept rather than failing the configure.

To produce a release zip, including the models:

cmake --build . --config Release --target zip -j

Creates build\dist\OLAS-win64-<version>.zip.

Cross-compiling from Linux with mingw-w64 is also supported via mingw-toolchain.cmake.

Open an issue at https://github.com/Igna-Mendez/Open-Local-Audio-Scribe-OLAS/issues with your Windows version and CPU model, the output of olas_win.exe -l en,es -v --stats, which mode you are running, and what you expected versus what happened.

MIT. Moonshine and miniaudio are both MIT; see LICENSE for the full text.