Murmur

Deutsch | English | Changelog

Murmur is a Windows dictation app. Hold your push-to-talk key, speak, then release it. Murmur formats the transcript and types it into the focused application.

The Math profile accepts spoken numbers and math expressions in German or English. The interface language does not affect which language it accepts.

Murmur is a separate, AI-assisted Windows adaptation inspired by per-simmons/murmur-youtube.

Download and first run

Download Murmur 0.3.0 beta for Windows 11 x64:

• Setup EXE installs Murmur for the current user.
• Portable ZIP: extract it and run Murmur.App.exe; it does not install Murmur.
• build-info.json records the source commit and build details. SHA256SUMS.txt contains checksums for the downloads.

The installer is unsigned, so Windows SmartScreen may show a warning. Murmur opens its setup screen on first run. Choose a microphone, hotkey, and model, then download the selected speech model. After setup, hold Right Ctrl to speak and release it to insert the text. You can change this key in Settings. Murmur stays in the system tray; open Settings or quit from its tray menu.

Requirements for release downloads

• Windows 11 x64 and a working microphone; allow desktop apps to access the microphone in Windows privacy settings.
• The Setup EXE and portable ZIP include the .NET runtime, UI libraries, and native speech engine. You do not need to install the .NET SDK, Python, or CUDA. Recognition runs on the CPU; a dedicated GPU is not required.
• Internet access to GitHub is required for the first model download inside Murmur. Model weights are not bundled with the installer. After the model is downloaded, speech recognition runs locally and works without an internet connection; the startup update check still contacts GitHub.
• Keep all extracted files together when using the portable ZIP. The application EXE depends on the libraries beside it.

You only need the .NET SDK if you build from source. Murmur extracts models with Windows 11's built-in tar.exe.

Murmur checks for a newer stable release at startup. If it finds one, the tray menu offers an action that opens that release page. It does not update itself. Beta releases are excluded from this check; check the Releases page manually for beta updates.

The model archives are about 465 MiB (Parakeet), 453 MiB (Nemotron), and 538 MiB (Whisper Turbo). Their extracted files use about 640 MiB, 653 MiB, and 989 MiB in the local release measurement. Allow extra free disk space for downloading and extraction. RAM use and speed depend on the model, audio session, and PC; there is no guaranteed RAM minimum or speed for every device.

Data locations

Speech and transcripts stay on the PC; they are not uploaded to a speech service. History is disabled by default on fresh installs. Stored settings, history, dictionary, and logs are local and unencrypted; review them before sharing. Logs are limited to a 2 MiB active file and two backups.

Path	Contents
%LOCALAPPDATA%\Murmur\settings.json	settings
%LOCALAPPDATA%\Murmur\history.jsonl	dictation history
%LOCALAPPDATA%\Murmur\dictionary.txt	personal dictionary
%LOCALAPPDATA%\Murmur\murmur.log	application log
%LOCALAPPDATA%\Murmur\models\	downloaded speech models
%LOCALAPPDATA%\Programs\Murmur	Setup EXE installation

What it does

• Captures microphone audio while your configured push-to-talk key is held.
• Runs Parakeet v3 or Whisper Large-v3-Turbo for final transcripts after key release, or Nemotron 3.5 for live transcript previews in the recording overlay. One model is resident at a time.
• Offers Standard and Math profiles. F8 switches profiles by default; you can change that key or switch through Settings and the tray menu.
• Converts supported German and English number words and math expressions into digits and mathematical symbols in the Math profile. Choose Unicode or LaTeX output.
• Performs conversational self-correction and speech repair (filtering slips of the tongue, filler words, and repair markers like "nicht X sondern Y", "ich meinte", "scratch that") with a configurable toggle in Settings.
• Restores focus to the target window before typing if the active window changed during recording, preventing accidental text insertion into other applications.
• Applies punctuation and optional filler removal, with a separate filler setting for each profile. Your personal dictionary supplies text corrections before insertion; edit or disable its entries in Settings.
• Types at the current cursor or the end of the document. If insertion fails after transcription, Settings offers the finished text for copying. This recovery text is temporary and does not survive an app restart.
• Keeps an optional dictation history with copying and deletion of individual entries or the whole history. Choose 7, 30, or 90 days of retention, or no age limit, and a 1,000-entry limit or no entry limit. History is disabled by default.
• Shows a recording overlay without taking focus from the target application; its screen position is configurable.
• Lets you select and test a microphone, choose an English or German interface, and enable startup with Windows.

All three speech models (Parakeet v3, Nemotron 3.5, and Whisper Large-v3-Turbo) support German and English among their languages. Recognition detects the spoken language automatically. The Math profile accepts either language, regardless of the interface language.

Math dictation

The Math profile accepts supported German and English number words and math expressions. English whole-number words are supported from 0 through 999; larger values can be dictated as digits. Unicode keeps German decimal commas and English decimal points. LaTeX uses a decimal point for both.

Spoken phrase	Unicode output
five times three equals fifteen	5 × 3 = 15.
x equals five point zero five	x = 5.05.
x to the power of two	x².
square root of x	√x.
x gleich fünf komma null fünf	x = 5,05.

In LaTeX mode, s squared plus square root of x becomes s^2 + \sqrt{x}. F8 switches between Standard and Math profiles by default. Formatting uses fixed rules; it does not solve equations or rewrite text with a language model.

Optional audio feedback

To play a sound after text is inserted, place your own WAV at Sounds\feedback.wav beside Murmur.App.exe and enable audio feedback in Settings. Murmur includes no sound files and stays silent if that file is missing.

Murmur uses .NET 10 and Avalonia. It captures audio with WASAPI via NAudio and runs speech models through sherpa-onnx.

Limits

Transcription accuracy has not been measured. The default model was not chosen from an accuracy benchmark.

Automated checks cover only part of the app's behavior. Compatibility with every microphone, PC, or target application has not been verified. Windows blocks text insertion into applications with higher privileges than Murmur.

License

Murmur source code is MIT licensed. Model and bundled third-party notices are in NOTICE.md.

Build

cd windows
dotnet restore Murmur.sln
dotnet build Murmur.sln --configuration Release --no-incremental -warnaserror
dotnet test Murmur.sln --configuration Release --no-build --filter 'Category!=RealModel&Category!=QualityCorpus'

This command runs the unit and UI tests used by normal CI. The release workflow also downloads models and runs inference tests marked RealModel. .github/workflows/real-model.yml runs model tests manually or nightly. Tests marked QualityCorpus require a separately supplied corpus.

More detail: Windows setup and build notes.