Local speech-to-text for Windows

Dictate. Record meetings.
All on your PC.

Hold a hotkey and your words land at the cursor, in any app. Press another and a call becomes a transcript with speakers. No cloud, no account. Audio and transcripts stay on your PC, unless you point NexusVox at a server of your own.

Get NexusVox See what it does ›

to transcribe 10 s of English speech on a GPU: 0.17 s; NVIDIA GPU, one test PC
to transcribe a 30-minute recording on a CPU: 100 s; Ryzen 7 5800X, 16 threads, no GPU
for private use: €0

Two jobs.
One local app.

Audio and transcripts never leave your PC. The exceptions are yours to set: a transcription, language-model or translator server that runs on another computer.

Dictation

Hold a hotkey, talk, let go. Your words land at the cursor, in any app.

Push to talk or toggle, with your own keys
Voice commands, dictionary, paste-last-text
Rewrite selected text by voice

Ctrl Win

Illustration Hold, speak, let go.

See dictation ›

Meetings

Record a call without a bot in it. One transcript, with numbered speakers.

Stop from anywhere with Ctrl+Shift+F9
Or upload a recording, with speaker detection
Export as TXT, SRT or VTT

Ctrl Shift F9

Illustration The meeting hotkey.

See meetings ›

Local.
Private.
Fast.

Where does your voice go?

Audio stays on your PC with the built-in models
No account, free for private use
Works offline once the model is downloaded

Every sentence is uploaded as audio
Needs an account, often a subscription
No internet, no dictation

Local

Runs on your machine.

The speech model runs on your NVIDIA GPU in Docker, or in-process on the CPU. Dictation and meetings alike. Nothing to sign up for.

Private

Your voice stays home.

Audio and transcripts never leave your PC. Transcripts live in a local SQLite file, audio as WAV files next to it. Turn audio storage off and only the text is kept.

Fast

0.17 s

to transcribe 10 s of English speech on a GPU. Parakeet TDT, 5.8 % word error rate. NVIDIA GPU, one test PC.

Audio goes nowhere with the built-in models. If you choose a transcription, language-model or translator server that runs on another computer, what you send to it goes there.

Measured, not promised.

Each figure comes from a test in the NexusVox repository, run on one PC. The details are on the features page.

to transcribe 10 s of English speech on a GPU: 0.17 s; NVIDIA GPU, one test PC
for a 3 s dictation on a CPU (median): 0.18 s; Ryzen 7 5800X, 16 threads
to transcribe a 30-minute recording on a CPU: 100 s; Ryzen 7 5800X, 16 threads, no GPU
word error rate, English, on a GPU: 5.8 %; FLEURS read speech, NVIDIA GPU

How we measured › All features ›

Up and running
in three commands.

Windows 11, Python 3.11 or newer, Docker for the GPU path.

PowerShell

            # 1. Get the code
            

            {s.command}
            {i < steps.length - 1 ? '
' : ''}
          }
            # 2. GPU only: start the speech model
            

            {s.command}
            {i < steps.length - 1 ? '
' : ''}
          }
            # 3. Install and run
            

            {s.command}
            {i < steps.length - 1 ? '
' : ''}
          }

No NVIDIA GPU? Use “Copy for CPU”: it leaves out step 2, and Parakeet runs on your CPU without Docker.

View on GitHub Full setup guide ›

Dictate. Record meetings.All on your PC.

Two jobs.One local app.