Record, transcribe and summarize your meetings without anything leaving your computer.
Features · How it works · Getting started · Privacy · Architecture · Building · Structure · Credits
LoopbackTranscriber is a Windows desktop application. It records the audio of one specific application (Teams, Zoom, Meet in the browser…) together with your microphone, transcribes it with Whisper, tells what you say apart from what the others say and, when you ask for it, writes a summary with Qwen3-4B. While you record, it can also show the meeting translated live in a floating window.
Everything runs on your computer. The internet is only used to download the two models.
|
Per-application recording Only the audio of the chosen program, plus your microphone on a separate track. If the application track turns out silent, the loudest default audio output is used instead. |
Any application Teams is built in. Zoom, Discord or the browser (for Meet) can be added from the processes that are playing audio. |
|
Transcription with Whisper Whisper large-v3-turbo, on a Vulkan GPU when there is one. Fixed language or automatic detection. Hallucinations, repetitions and echo are filtered out.
|
Who is speaking Your microphone is transcribed separately and aligned with the meeting. Each paragraph is labelled Participants or You; if you have a name set (by default, your Windows first name), it reads pgomur (You). |
|
History and search Search the title, application, date, summary and full text, ignoring case and accents. Filter by date and duration, and transcribe again with the current settings. |
AI summary Qwen3-4B writes a title, purpose, key points, topics, goals and next steps, in the application language. |
|
Live translation A floating window with what the participants say and its translation into Spanish, English, Portuguese, German or Russian. Without the summary model, it only transcribes. |
Nothing is lost After an error, an exit, a shutdown or a crash, the audio is kept as an unprocessed recording so it can be retried. |
The interface, the meeting language and the translation are available in Spanish, English, Portuguese, German and Russian. The meeting language can also be detected automatically.
- Capture. The application is recorded with Windows process loopback, so only its audio (and that of its child processes) is captured. The default audio outputs (general and communications) and the microphone are recorded at the same time on separate tracks.
- Preparation. The tracks are converted to 16 kHz mono, and a noise gate silences the microphone between your words.
- Transcription. Whisper transcribes the meeting and your microphone, each in its own worker process: in parallel, or one after the other when the language is detected automatically. Each process exits when done to free the native and GPU memory.
- Composition. Your microphone is aligned with the meeting, sentences picked up by both tracks (speakers into the mic, or your voice played back by the application) are kept only once, and each paragraph gets its speaker label.
- Summary. On demand, Qwen3-4B summarizes the transcript, also in its own worker process.
| Item | Requirement |
|---|---|
| System | 64-bit Windows 11 (10.0.22000 or later) |
| Acceleration | A Vulkan GPU when there is one; otherwise the CPU (slower) |
| Runtime | Microsoft Visual C++ Redistributable (x64), added by the installer if it is missing |
| Disk | About 3.1 GB for the two models |
| Model | Used for | Size |
|---|---|---|
| Whisper large-v3-turbo (GGML, q5_1) | Transcription and live transcription | ~624 MB |
| Qwen3-4B (GGUF, Q4_K_M) | Summaries and live translation | ~2.5 GB |
- Download
LoopbackTranscriberSetup-<version>.exefrom the latest release and run it. The wizard is available in Spanish, English, Portuguese, German and Russian. The installer is not code-signed: if Windows SmartScreen shows "Windows protected your PC", click More info → Run anyway. - On the first start, the Before you start screen downloads both models. Each download can be paused, resumed or cancelled, and its SHA256 checksum is verified when it finishes.
- In Settings, check your name, the meeting language and the application language.
Only one instance can run at a time: opening it again shows the window of the one already running.
- Pick the application. Teams is built in; others are added in Applications.
- Record. Press Start recording, or use the tray icon.
- Get the transcript. Press Stop recording. Whisper transcribes it and, if your microphone was recorded, separates Participants from You. It is saved in the history.
- Summarize. In the AI summary tab, press Summarize.
The window's X does not close the application: it keeps running in the tray, so a recording can continue. To exit, right-click the tray icon and choose Exit.
Optionally, open live translation to follow the meeting translated while you record. Full walkthrough in the user guide, and common problems in Troubleshooting.
Audio, transcripts and summaries never leave your computer. The application only goes online to download the models, from a fixed Hugging Face revision and with their SHA256 checked. The installer may also download the Visual C++ Redistributable if it is missing.
%LocalAppData%\LoopbackTranscriber\
├── settings.json
├── models\ Whisper and Qwen3-4B (configurable)
├── transcripts\ One folder per application and per recording (configurable)
├── failed\ Unprocessed recordings, kept to be retried
├── logs\app.log
└── live-metrics.log Live translation timings
Each recording is a folder transcripts\<App>\<App>_<yyyyMMdd_HHmmss>\:
| File | Contents | When it exists |
|---|---|---|
<base>.md |
The transcript, in Markdown | Always |
<base>.wav |
The main audio track | Always |
<base>_mic.wav |
Your microphone track | If the microphone was recorded |
<base>.recording.json |
The microphone offset | If there is a _mic.wav |
<base>.summary.json |
The summary | If one was generated |
When uninstalling, the wizard asks whether to also delete %LocalAppData%\LoopbackTranscriber (the default answer is No). Models or transcripts folders moved elsewhere in Settings are not touched.
Important
Only record and transcribe with the consent of the people taking part. You are responsible for informing them and for complying with the law in your country.
A single WPF project following MVVM with CommunityToolkit.Mvvm and a Microsoft.Extensions.DependencyInjection container.
| Layer | Contents |
|---|---|
| Views | XAML and code-behind: main window, tray menu, floating translation window |
| ViewModels | State and commands of each screen; they talk to each other through messages |
| Services | Capture, processing, transcription, summary, live session, persistence, localization and logging |
| Interop | WASAPI COM interfaces, process loopback activation, audio sessions and the process tree |
Design principles:
- Nothing leaves the computer. The only network access is downloading the models.
- No audio is lost. A recording that cannot be transcribed is kept to be retried, even if the application closes, hangs or the computer shuts down.
- Inference runs in separate processes. Each transcription and each summary runs in a child process, so the application stays light while it waits in the tray.
- Services do not depend on WPF. The system dependencies sit behind interfaces, so the logic is tested without a window, audio devices or models.
| Package | Version | Used for |
|---|---|---|
| CommunityToolkit.Mvvm | 8.4.2 | Observable properties, commands and messaging |
| Microsoft.Extensions.DependencyInjection | 10.0.0 | Service container |
| WPF-UI | 4.3.0 | Dark theme, accent color and Fluent controls |
| NAudio | 2.2.1 | Device and microphone capture; WAV reading, writing and conversion |
| Whisper.net (+ Runtime, Runtime.Vulkan) | 1.9.1 | Speech recognition on the CPU or on a Vulkan GPU |
| LLamaSharp (+ Backend.Cpu, Backend.Vulkan) | 0.27.0 | Summaries and translation with Qwen3-4B |
More in Architecture.
Requires 64-bit Windows 11 and the .NET 10 SDK. Visual Studio is optional.
dotnet build LoopbackTranscriber.slnx -c Release
dotnet test LoopbackTranscriber.slnx -c Release
dotnet publish LoopbackTranscriber\LoopbackTranscriber.csproj -p:PublishProfile=FolderProfilePublishing writes a self-contained win-x64 build to installer\publish\. The installer is compiled with Inno Setup 6.3 or later and is written to installer\Output\LoopbackTranscriberSetup-<version>.exe:
ISCC.exe installer\LoopbackTranscriber.iss| Topic | Detail |
|---|---|
| Version | Defined only in Directory.Build.props; the installer reads it from the published executable |
| Quality | .NET analyzers at latest-recommended, and warnings treated as errors |
| Localization | Every build checks the interface texts: an unknown key or a missing translation is an error |
| Tests | xUnit, split into Unit and Integration: dotnet test LoopbackTranscriber.slnx --filter Category=Unit |
| Documentation | DocFX site: docs\build.ps1 builds it, docs\publish.ps1 -Push publishes it |
More in Building and publishing, Tests and Installer.
loopback-transcriber/
├── LoopbackTranscriber/ The application (WPF, net10.0-windows)
│ ├── Program.cs, App.xaml Entry point and startup
│ ├── AppServices.cs Service container
│ ├── MainWindow.xaml Main window
│ ├── Views/ Screens, tray menu and floating window
│ ├── Converters/, Styles/ Binding converters and shared styles
│ ├── ViewModels/ State and commands of each screen
│ ├── Services/ Capture, processing, summary, live session…
│ ├── Models/ Plain data
│ ├── Interop/ WASAPI and process loopback
│ ├── Assets/ Application icon
│ ├── Properties/ Publish profile
│ └── Build/ Build-time translation check
├── LoopbackTranscriber.Tests/ xUnit tests with test doubles
├── installer/ Inno Setup script
├── docs/ DocFX documentation (English and Spanish)
├── Directory.Build.props Version and analyzers
└── LoopbackTranscriber.slnx Solution
Built with CommunityToolkit.Mvvm, WPF-UI, NAudio, Whisper.net (whisper.cpp), LLamaSharp (llama.cpp) and Microsoft.Extensions.DependencyInjection.
The models are not included in the installer: OpenAI Whisper (MIT) and Qwen3-4B (Apache 2.0). Full notices in THIRD-PARTY-NOTICES.txt.

