Dictation

Speak.
It's typed.

Hold a hotkey, talk, let go. Your words land at the cursor, in any app. On a GPU or just the CPU, no cloud, no account.

Get NexusVox Also: record meetings ›
Meeting notes.txt — Notepad

Sprint review, Oct 17

Demo went well, two open bugs on the export path.

Let's move the launch review to Thursday at 3 pm and invite the design team.

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
word error rate, English, on a GPU
5.8 %
FLEURS read speech, NVIDIA GPU

Measured on one PC. Details and sources ›

Everything dictation does

Push to talk

Hold. Talk. Let go.

Hold Ctrl+Win, speak, release: the text lands at the cursor. Prefer one press to start and one to stop? Switch to toggle mode. Bind any key combination with one click.

Illustration Hold to talk, let go to type. Default keys; change them in Settings.

Waveform overlay

See that it listens.

A small waveform at the bottom of the screen shows that NexusVox records. It never takes focus and ignores the mouse. Prefer sound? Pick chimes, the waveform, or both.

Illustration Illustration of the waveform. It shows only while recording.

Voice commands

Say “new paragraph”. Get one.

Spoken shortcuts for line breaks, tabs, capital letters and punctuation, plus number words as digits. English and German. An optional “press enter” at the end sends a chat message.

You sayIt types

“new paragraph” a blank line

“tab” a tab

“all caps” the next words in capitals

“twenty three” 23

Illustration Examples of spoken shortcuts.

Paste last text

Say it again.

Hold the hotkey, tap V, let go: nothing is transcribed, the text NexusVox pasted last is pasted once more.

Illustration Hold the dictation keys, tap V, let go. The paste key is configurable.

Polish

Select. Speak. Rewritten.

Select text, hold the hotkey, tap the polish key and say what to do: “make it shorter”, “mach es formeller”. A local language model rewrites the selection in place, and Ctrl+Z undoes it.

  • Needs a model server you run: Ollama by default, or any OpenAI-compatible server. If that server is on another computer, your selected text goes to it.
  • Off until you pick a polish key.
to rewrite a short selection on a GPU
0.54 s
RTX 5070 Ti, gemma4:e4b
to rewrite a short selection on a CPU
2.0 s
Ryzen 7 5800X, 8 cores, gemma4:e2b
Time to rewrite a short selection, with the model loaded.

Polish presets

Name it once. Say the name.

Keep instructions you use often under a short name, like “prompt engineer” or “formelle Mail”. Say the name instead of the whole instruction, add words after it to adjust it, and mark one as the default for a tap of the polish key without a word.

  1. Save an instruction under a name
  2. Say the name
  3. The selection is rewritten
Illustration How a preset works.

Auto cleanup

Talk like you talk. Get clean text.

Removes fillers (“um”, “äh”), false starts and repeated words, and fixes punctuation. The medium level fixes grammar too. If the reply does not look like your words cleaned up, NexusVox types what you said.

  • Off by default. Uses the same language-model server as polish.
  • Without an NVIDIA GPU, NexusVox picks the smaller gemma4:e2b by default; the figures use it on the CPU and gemma4:e4b on the GPU.
for Auto cleanup of a sentence on a GPU
0.53 s
RTX 5070 Ti, gemma4:e4b
for Auto cleanup of a sentence on a CPU
2.9 s
Ryzen 7 5800X, 8 cores, gemma4:e2b

You sayIt types

“um, so I wanted, I wanted to say thanks” So I wanted to say thanks.

Illustration Illustration of a light cleanup, not a measured result.

Correction dictionary

Teach it your words.

Names, product terms, jargon: fix a word once and every later transcript gets it right.

HeardTyped after you fixed it once

“next o box” NexusVox

Illustration Illustration: one dictionary entry.

Learn from edits

Fix it once. It remembers.

Correct a misheard word right in the app you dictated into, and NexusVox suggests a dictionary entry for it. You accept or dismiss it in Settings. It looks at the text field for 20 seconds after a paste, and never at password fields or terminals.

  • Off by default.
  1. Dictate into any app
  2. Fix a misheard word there
  3. Accept the suggested dictionary entry
Illustration How learn from edits works.

Review and retry

Switched models? Try again.

Replay a recording in the Review tab. Retry transcribes its stored audio again with the model you use now.

  • Needs “Store audio recordings” turned on.
  1. Replay a recording
  2. Press Retry
  3. Get a new transcript from the current model
Illustration How retry works.

Dashboard

Every word, counted.

Words per minute, minutes saved, usage over time. Every transcription is stored on your PC and analysed there.

Dashboard, Analytics tab. Demo data, not a real user.
Dashboard, Analytics tab. Demo data, not a real user.

What the numbers mean in practice

Short messages

Example

A 3 s dictation is transcribed in 0.18 s on a desktop CPU.

Reply to a chat, fill a form, jot a note. That is the median on one fast desktop CPU with 16 threads; the text still has to be typed after that. A slower CPU takes proportionally longer.

for a 3 s dictation on a CPU (median)
0.18 s
Ryzen 7 5800X, 16 threads
for an 8 s dictation on a CPU (median)
0.34 s
Ryzen 7 5800X, 16 threads
for a 3 s dictation on a GPU (median)
0.08 s
NVIDIA RTX 5070 Ti
0.18 s CPU only, 3 s dictation
→
0.08 s With an NVIDIA GPU

An example from our own measurements, not a customer story.

Long dictation

Example

A minute of speech in one go scores about the same as short clips.

Dictate a paragraph or a whole email without stopping. On the CPU, long German speech and short German clips land within a fraction of a point of each other. The test material is read speech, which is easier than a noisy room.

word error rate, English, 48 to 55 s of continuous speech on a CPU
2.4 %
audiobook reading, CPU
word error rate, German, items of about 57 s on a CPU
7.1 %
FLEURS clips joined, CPU
word error rate, German, short clips on a CPU
7.3 %
FLEURS read speech, CPU
7.3 % German, short clips
→
7.1 % German, about 57 s

An example from our own measurements, not a customer story.

No graphics card

Example

No graphics card needed. The CPU sets the speed.

Parakeet runs on the CPU, with no Docker. These are one desktop CPU pinned to fewer threads; a slower CPU will be slower still. Eight threads is the recommended minimum, four threads work but you wait.

for an 8 s dictation on a CPU (median)
0.34 s
Ryzen 7 5800X, 16 threads
for an 8 s dictation, pinned to 8 threads
0.51 s
Ryzen 7 5800X, 8 threads
for an 8 s dictation, pinned to 4 threads
2.5 s
Ryzen 7 5800X, 4 threads
2.5 s 8 s dictation, 4 threads
→
0.34 s Same dictation, 16 threads

An example from our own measurements, not a customer story.

Ready to try it?

One download for dictation and meetings. Windows 11, free for private use.

Get NexusVox Also: record meetings ›