Talk through the commit. Paste the message.
You already explain the change out loud - in standup, in review, to yourself. Voxmelt catches that and hands back the written version, without your source code going anywhere near a cloud model.
Transcription runs on your CPU. The rewrite runs on a model on your own GPU via Ollama. No API key exists because there is no API to key into. Pull the network cable and it behaves identically - which is the only version of this claim worth making, because you can check it in a minute with any network monitor.
What you mumble, and what lands at the cursor
These are real template outputs, not aspirational ones. The messy input is the point - fillers, restarts and thinking out loud are allowed.
It does not want your GPU
Dictation uses zero VRAM
Speech-to-text runs on CPU cores. Your card stays free for the build, the container, the training run or the game.
No API key, no rate limit
Nothing to provision, nothing to rotate, nothing to expense. It works the first time you press the hotkey.
Works air-gapped
After the one-time model download it needs no network at all. Same behaviour on a plane or a locked-down box.
Any app, one hotkey
Editor, terminal, browser, Slack. No plugin per tool, no per-app configuration.
Where it is weakest, since you will find out anyway
Code is the hardest clip in our benchmark: 11.9% word error overall, 20.7% on the code clip. Most of that is tokenisation rather than mishearing - identifiers come back as separate words - and custom vocabulary fixes the ones you use daily. It is also Windows only, and the AI rewriting needs an NVIDIA card. The full numbers, including the clips we lose on, are on the benchmark page, with a harness you can download and rerun on your own audio.
What developers ask first
Does my source code ever leave the machine?
No. Speech-to-text runs on your CPU and the AI rewriting runs on a model on your own GPU through Ollama. Neither sends audio or text anywhere. The only outbound calls Voxmelt makes at all are model downloads, a licence check, and update checks - so you can dictate about a private repository with the network cable out and it behaves identically. Verify it the way you would verify anything: run a network monitor while you dictate.
How well does it handle identifiers, CamelCase and technical terms?
Honestly: this is the weakest clip in our published benchmark. Parakeet scores 11.9 percent word error overall but 20.7 percent on the code clip, and most of that gap is tokenisation rather than mishearing - "useDebounce" comes back as "use debounce". Custom vocabulary (Pro and Studio) is the fix: teach it your project's names once and it corrects them on every engine afterwards. Do not expect it to spell an unfamiliar API surface correctly out of the box.
Can I dictate directly into VS Code or a terminal?
Yes. A global hotkey works from any Windows app, and the polished text is on your clipboard - with auto-paste it lands at the cursor. Mini Mode is a floating pill that stays on top, so you can dictate into an editor without switching windows. There is no editor plugin to install and nothing to configure per app.
Do I need an NVIDIA GPU?
Not for dictation - that runs on any modern CPU and uses no VRAM at all, which matters if your card is already busy with a build, a container, or a game. The AI rewriting does need an NVIDIA GPU with 6 GB or more, because that is where the local model runs. Without one you still get unlimited local transcription; you just paste the raw transcript instead of the polished version.
Is there an API or CLI?
No. Voxmelt is a desktop app with a global hotkey, not a service you script against. If you want a local transcription API, faster-whisper and whisper.cpp are the right tools and they are free - we would rather say so than sell you the wrong thing.
What does it cost?
Free tier is 60 minutes of dictation a day, permanently, with no card. Pro is $49 once and Studio $99 once - one payment, no renewal, nothing to cancel. Every install starts with 15 days of the full Studio feature set so you can judge it on your own hardware first.
Try it on your own repo
15 days of everything, no card. After that the free tier keeps going at 60 minutes a day, and $49 once buys Pro outright.