no cloud processing · no graphics card required · a GPU only adds more

Dictation that runs on the PC you already have.

No graphics card? It works, fully. Have an NVIDIA card? You also get the most accurate engine we ship and the local AI studio.

Press a hotkey and speak. NVIDIA's Parakeet v3 transcribes on your CPU, so the no-GPU path is the whole product rather than a trial mode: 11.9% word error across our five published clips, and 2.0% on the fast natural speech clip, where Whisper large-v3 on an RTX 3080 Ti scores 3.1%. Add an NVIDIA card and you get large-v3 itself, still the most accurate overall at 7.8%, plus a local model that rewrites and translates. Your audio and transcripts never leave the machine.

Press the hotkeyin any app
Speakstreaming · uncapped on Pro and Studio
It transcribesCPU by default · a GPU instead
AI polishes80+ tones · local · optional
Text appearsin Voxmelt · ready to copy

Real dictation on a real machine, no cuts. Everything you see runs on the CPU; the network is doing nothing.

What we store, plainly

No hedging, and nothing here needs taking on trust - run any network monitor while you dictate.

Your audioNever leaves the machine. We never receive it, so we could not store it if we wanted to.
Your transcriptsLocal only. You choose whether they are kept at all - History has a "don’t save" mode.
Your accountEmail, licence state and, if you buy, your payment record, on our server, plus a hashed device ID that anchors the trial, logged with the IP address and user agent it arrives with. We never see your card.
TelemetryThe app has no usage analytics and no crash reporting; this website keeps cookieless page statistics. See everything the app sends.

Voxmelt goes online only for the trial and licence check (a hashed device ID), the optional account and purchase, and anonymous update and model downloads. None of it carries your audio or text. Full privacy policy

Voxmelt
Studio · 9 days left Sign in
COLOR THEME
Nightwing
Matrix
Cyberpunk
Candy Red
Mono
RecordWhisper small→Clean up · Default
AI TONE / PRESET
Clean up · Default
Coding prompt · Feature
Chat Responder · Corporate Polish
Summarize · Bullets
Custom prompt · Editable
→gemma2:27b
AI MODEL · OLLAMA
gemma2:27b · 14.56 GB
gemma3:4b · 3.11 GB
gemma4:e4b · 8.95 GB
qwen3:8b · 5.20 GB
qwen3:14b · 9.20 GB
Free GPU Voice Off AI On
RECORD
COMPOSE
VOICEOVER
TEMPLATES
HISTORY
MINI
NVIDIA RTX
GPU
20%
VRAM
4.5/12G
TEMP
42°C
PWR
138W
Click to record
Ctrl+Shift+R
Microphone access
Whisper small loaded
GPU available
TRANSCRIPT0 words
ProcessAC
Paste or type text here to process with AI, or use the mic to record…
AI OUTPUT
AC
AI output will appear here
Record or paste text, then process
MODELS
Whisper small loadedgemma2:27b off
WHISPERsmall
Tiny↓75 MB ~1.2 GB
Base↓145 MB ~1.3 GB
Small↓480 MB ~1.6 GB
Medium↓1.5 GB ~2.6 GB
Large v3↓3.0 GB ~4.2 GB

Small is the active model. great accuracy · uses ~1.6 GB VRAM.

Load on app launch
Pre-loads the model so the first recording is instant.
Advanced
AI PROCESSINGgemma2:27b
Keep both loaded
Both share VRAM · needs 24+ GB

Performance may be unstable. Both models compete for VRAM.

Connected · 5 models installedRefresh
Ollama v0.24.0INSTALLED MODELS
gemma2:27b
14.56 GB on disk
Remove
gemma3:4b
3.11 GB on disk
Remove
gemma4:e4b
8.95 GB on disk
Remove
qwen3:8b
5.20 GB on disk
Remove
qwen3:14b
9.20 GB on disk
Remove
Readysmall / float16NVIDIA RTX
Ctrl+Shift+R to toggle
DISPLAY 1 · 27″
2560 × 1440 · the workbench
prompt.txt
FileEditView
Paste your cleaned coder prompt here…
Ln 1, Col 1UTF-8100%
READY - TAP TO DICTATE
Tap mic to dictate
TRANSCRIPT AI
Your transcript will appear here as you speak…
DISPLAY 2 · 24″ portrait
1080 × 1920 · Notepad
Audio stays local No API keys Always-on-top Real-time sync Voice commands 80+ AI tones Custom prompts 1-tap copy Per-app paste Whisper large-v3 Ollama-powered Works offline after setup 11-13x realtime Auto VRAM swap Global hotkey 5 themes Export to WAV (Pro) Compose studio
why people melt local

No cloud processing - no audio upload, no per-minute fees.

voice on your processor · graphics card reserved for AI
PARAKEET V3CPU STT ENGINE
VX·CORELLM RUNTIME
voice in
AI output · streaming
cleanedpunctuatedon-device
11-13xrealtime STT
CPU LOAD34%
Parakeet v3STT
VRAM6.2 / 12 GB
local LLMcleanup
TEMP61°C

The engines are already warm - Parakeet on your CPU, your AI model on the GPU. While they sit loaded, you get four creative surfaces at no extra cost:

Mini Mode

Always-on-top floating widget. Dictate over any app without switching windows - one tap to record, the transcript lands at your cursor.

Compose

Paste any text, pick a template and tone, the local LLM reshapes it. Emails, tickets, commit messages - no mic needed, same GPU.

Voiceover

Local text-to-speech, runs on your CPU. 4 personas x 8 tones = 32 voice combinations. Export WAV (Pro). No per-character billing, no cloud TTS.

Hands-free Mode

Voice commands start and stop recording, copy or clear the text, and load or unload models - without touching the keyboard.

Windows 10 (64-bit) and 11 on x64 PCs · NVIDIA GPU optionalaudio never leaves the device10 min of speech, typed in ~50 secondsEN + 24 European languages on CPU · ~100 via WhisperParakeet dictation + AI studio + voiceover · nobody else bundles thisbuilt for the personal AI computer era
new · May 31, 2026NVIDIA + Microsoft · RTX Spark keynote

The PC just got reinvented around local, private AI on NVIDIA. NVIDIA and Microsoft unveiled RTX Spark - a new class of Windows PC built to run AI on your device, not the cloud. Voxmelt has done exactly that for your voice since day one. You don't need to wait for fall hardware to live in that future - dictation already runs on the CPU you own, and the NVIDIA GPU in your rig powers the local AI studio.

Voxmelt is an independent product, not affiliated with or endorsed by NVIDIA or Microsoft.

Published in August 2026 and recorded on an earlier GPU-first setup: install, both engines, AI cleanup, templates and voiceover. Today dictation needs no graphics card; the AI studio still needs an NVIDIA GPU. Nothing loads from YouTube until you press play.

the thermal pipeline

Two heavy AI models. One GPU. Fully orchestrated.

The all-GPU mode, for people who want it: flip dictation from the default CPU engine to Whisper on CUDA and Voxmelt runs the speech engine and the LLM cleanup engine on the same consumer GPU - both kept in VRAM when they fit, or swapped one at a time if you flip the switch. Tap a model to load it, or flip the dial to swap mode and watch only one ride the card at a time.

ENGINE 01
STT engine
Whisper large-v3
the recording brain
Throughput10-15x realtime
Precisionfloat16 · int8
VRAM~4.2 GB
Languages99 + auto-detect
BackendWhisper · CUDA
separate engine process · stdio, online only for model weights
GPU orchestrator
RTX 3080 Ti· 12 GB
VRAM
129630
0.0/12G
STTLLM
ENGINE 02
LLM engine
Ollama models, 3B to 70B
the cleanup brain
Models22+ catalogued
VRAM2.6 GB to 42 GB
Outputtoken-stream
Presets16 built-in + custom
BackendOllama · 127.0.0.1
localhost-only · no text sent off the PC
Auto memory math
Checks your card's VRAM before each AI run: keep both models loaded, or swap one in at a time, and a model that will not fit is refused.
Zero-touch handoff
You stop talking and, with AI on, polished text streams in by itself - token by token. No button to press.
Built for your GPU
12 GB card? Whisper and a 4B model stay warm with headroom. 24 GB? Both stay loaded with any model up to 14B. 48 GB? 70B models on Studio: flip one switch and Whisper steps aside while they run.
see it cook

Speak. Stop. It's already done.

No "go" button, no second-guessing. The instant your mic cools, the cleanup fires and polished text streams in token by token. Here's the whole cycle, faked for the browser - the real thing runs on your NVIDIA GPU, with no internet once the models are downloaded.

0.0s
Ready
Transcript · Whisper large-v3
Click the mic. We'll fake-record a sample, transcribe it, then clean it up.
AI Cleanup · Gemma 3 · 4B
Empty
Local · no uploadNVIDIA RTX · 12 GB
not for everyone

For people who keep their data on their own machine.

If you keep a hard rule that sensitive audio never leaves the building, this was built for you. Any Windows 10 (64-bit, build 19041 or later) or 11 PC with an x64 processor runs it - an NVIDIA GPU just adds the local AI studio.

Developers

Talk through a commit message or a spec; clean text appears in Voxmelt, ready to copy. Works offline after the model download.

Creators & streamers

Dictate scripts on your CPU while the RTX card stays free for editing and games. Voiceover reads them aloud and exports a WAV (Pro).

Writers

Draft at the speed of speech. A local LLM tidies grammar and filler so the first pass reads like a third.

Clinicians

Dictate notes between patients on your own PC; no audio file is kept and no server holds your audio or text. There is no Voxmelt-side processor, but your own privacy and legal duties still apply.

Legal & finance

Dictate client notes and letters on your own PC; no audio file is kept. There is no Voxmelt-side processor, but your own privacy and legal duties still apply.

Offline teams

Works with the network cable unplugged once models are downloaded; paid licences work offline for 30 days after their last check. Verify it with a network monitor.

what it actually does

Talk like you talk. Send what comes out.

Pick whoever sounds most like you. The messy version on the left is the point - fillers, restarts and thinking out loud are what it expects.

what you say

what you get
on-ramp to the rtx spark era

The future they announced. The hardware you already own.

Jensen Huang says “the PC is being reinvented.” Satya Nadella wants “unmetered intelligence to every home and every desk.” That future ships on RTX Spark Windows PCs. But you don't have to wait for new silicon: the RTX card in your rig runs the whole local pipeline today.

no per-minute cloud meter, by designprivate by architecture - audio never leaves your machinelocal AI on the NVIDIA GPU you already paid forno account needed to dictate - works offline after the one-time model download

Voxmelt is independent and not affiliated with, sponsored by, or endorsed by NVIDIA or Microsoft. “NVIDIA,” “RTX Spark,” and “Windows” are trademarks of their respective owners, used here descriptively only.

Melt local. Keep private.

Microsoft Store install with no Windows warning, or the 5 MB direct installer (Windows will warn). Either way, Voxmelt downloads about 840 MB before your first dictation. Dictation and voiceover run on any modern x64 CPU - an NVIDIA GPU adds the local AI text studio. 15-day free run of the full Studio pipeline - no card on file, no catch.

Direct installer, for PCs without the Store; Windows will warn because it is not code-signed.

Windows 10/11 · dictation on any modern x64 CPU · NVIDIA GPU optional for AI · audio and text stay on your PC