OpenWhispr on Windows, or Voxmelt? Free and open, or local only.
Choosing between a free, open-source app that runs local or cloud models and one that runs only on your own PC? Here is what each costs, where your audio goes by default on Windows, whether each works offline, what each sends and how each rewrites text, with every OpenWhispr fact linked to OpenWhispr’s own repository and the date we checked it.
Free, open and cross-platform, or local only by design
If you want free, open-source software, also use a Mac or Linux, or need meeting transcripts and audio files transcribed, OpenWhispr covers more and costs nothing; set it to local models and your audio stays on your device. If you want dictation and AI rewriting that only ever run on your Windows PC, with no cloud default to change, plus read-back, and you are happy to pay $49 once after a free trial, that is what Voxmelt is for.
Sources: OpenWhispr README, OpenWhispr on GitHub; MIT License, OpenWhispr on GitHub; Network Allowlist, OpenWhispr on GitHub. Checked on 27 October 2026.
OpenWhispr
Free and open source, local or cloud
macOS, Windows and Linux, free under the MIT licence. Local models keep your audio on your device; OpenWhispr Cloud, the default after onboarding, runs transcription and AI in its cloud. Adds meeting transcripts, audio file import and an AI assistant.
Sources: OpenWhispr README, OpenWhispr on GitHub; MIT License, OpenWhispr on GitHub; Network Allowlist, OpenWhispr on GitHub. Checked on 27 October 2026.
Voxmelt
Local only, on your Windows PC
Windows 10 and 11 only. Your speech is recognised on your own CPU or GPU and the AI rewrites on your own NVIDIA GPU, with no cloud mode, so your audio and transcripts never leave the PC. Free tier, then $49 or $99 once, with no subscription.
OpenWhispr costs nothing. Here is how to keep it local.
If OpenWhispr, set to local models, does what you need, you do not need Voxmelt. Two things to know before you install it.
OpenWhispr, free and open source
OpenWhispr calls itself the open-source and free alternative to WisprFlow and Granola, and its code is under the MIT licence, free for personal and commercial use. Installers are on its GitHub releases page for Windows, macOS and Linux.
Sources: OpenWhispr README, OpenWhispr on GitHub; MIT License, OpenWhispr on GitHub. Checked on 27 October 2026.
Keeping it on your PC
Its network notes list OpenWhispr Cloud as the default after onboarding, and say local models are downloaded only when you opt into one: Whisper, Parakeet, or a local reasoning model. With local models, its README says, your audio never leaves your device, and every core feature, AI reasoning included, can run on them.
Sources: Network Allowlist, OpenWhispr on GitHub; OpenWhispr README, OpenWhispr on GitHub. Checked on 27 October 2026.
Windows also has dictation of its own. What voice typing (Win+H) and voice access do, and what Microsoft says about where your audio goes, are on Voxmelt vs Windows voice access and Win+H.
What each one does on Windows, checked against OpenWhispr’s own repository
| Voxmelt | OpenWhispr | |
|---|---|---|
| Price | Free tier: a 15-day trial of everything, then 60 minutes of dictation a day, no card. Pro $49 or Studio $99, paid once; no subscription | The app is free. Its hosted service, OpenWhispr Cloud, meters a weekly allowance, and its changelog mentions paid plans and subscriptions[1][3] |
| Source code | Not open source: you get a licence to use the app | Open source under the MIT licence, free for personal and commercial use[5][1] |
| Windows and hardware | Windows 10 (64-bit, build 19041 or later) and Windows 11 on an x64 CPU, with 8 GB of RAM. No graphics card to dictate; an NVIDIA GPU with 6 GB or more (Ampere or newer) for the AI text studio | A Windows installer (.exe), included in release 1.10.4; the README states no Windows version or hardware minimum. Its local Whisper can use the GPU through CUDA, or through Vulkan on AMD and Intel[1][4] |
| Other platforms | None: Windows only | macOS (Apple Silicon and Intel) and Linux; its changelog also mentions a mobile app, with subscriptions bought in the App Store or Google Play[1][3] |
| Where your audio goes | Stays on your PC: recognised on your own CPU or GPU, with no cloud mode | Your choice of local or cloud models. OpenWhispr Cloud, the default after onboarding, runs transcription and AI reasoning through its cloud API; with local models your audio stays on your device[2][1] |
| Works offline | Yes, after the model download at setup; a paid licence keeps working offline for 30 days after its last successful check | Yes on local models, which its README calls fully private offline transcription; each one downloads when you opt into it, and cloud processing needs a connection[1][2] |
| Signing in | You can record, transcribe and use the AI features with no account; an optional account attaches a paid licence | OpenWhispr Cloud uses an account sign-in, and team spaces are for signed-in users; its network notes also describe signed-out, fully local setups[2][1] |
| Data and telemetry | No audio or text on a server. A hashed device fingerprint for the trial and licence check, your email only if you create the optional account, payment records if you buy, and anonymous update and model downloads; the app has no telemetry, analytics or crash reporting | No data collection and no telemetry, its README says. With OpenWhispr Cloud, transcription, sync, agent reasoning, settings and usage go through its cloud API[1][2] |
| Dictating into other apps | A global hotkey and voice commands: the text appears in Voxmelt, ready to copy; Mini Mode, a small always-on-top widget, pastes it at your cursor | A global hotkey to dictate into any app with automatic pasting, and a second hotkey that dictates in one language and pastes the text in another[1] |
| AI rewriting | On your PC only, through Ollama on an NVIDIA GPU: 17 templates across 82 tones, plus your own templates | Cleanup in the tone you pick for each kind of app (Tone & Style, off by default), and an AI agent on GPT-5, Claude, Gemini, Groq, Tinfoil, OpenRouter or local models[3][1] |
| Meetings and audio files | No: Voxmelt dictates from your microphone and keeps no audio file | Detects Zoom, Teams and FaceTime calls and transcribes them with speaker labels; imports audio and video files and YouTube links[1] |
| Read-back | On-device text-to-speech reads text back, offline once the voice files are downloaded; WAV export on Pro and Studio | Not among the features its README lists[1] |
| Installs and updates | From the Microsoft Store, which keeps it updated, or a direct installer that checks each update against our signing key | An installer from its GitHub releases, 1.10.4 the latest, dated 21 October 2026 in its changelog; the app updates itself from GitHub, its network notes say[4][3][2] |
| Published accuracy | 7.8% word error rate with Whisper large-v3 on an NVIDIA GeForce RTX 3080 Ti (12 GB), across 5 recorded clips, published 16 July 2026 | None in its README[1] |
Each bracketed number links to OpenWhispr’s page for that fact; the full list is under Sources. Checked on 27 October 2026. Voxmelt’s accuracy comes from our published benchmark, which did not test OpenWhispr, its prices from the pricing page, its requirements from the download page, and what can leave your PC from our privacy policy.
When OpenWhispr is the better pick
If one of these is you, use OpenWhispr. We would rather lose the sale than the argument.
What you get with Voxmelt
No cloud mode at all
Speech is recognised on your own CPU or GPU and the AI rewrites on your own NVIDIA GPU, so your audio and transcripts never leave the PC, whichever template you use, and there is no cloud default to switch off. What does go online is the trial and licence check, the optional account if you make one, billing if you buy, update checks and model downloads, and none of it carries your audio or text.
Dictation and AI with no account
You can record, transcribe and use the AI features with no account. You make one only to attach a paid licence, and the app has no telemetry, analytics or crash reporting.
Dictation with the Wi-Fi off
Once its models are downloaded, Voxmelt dictates with no internet connection, and its AI rewriting runs offline on your own NVIDIA GPU too. A paid licence keeps working offline for 30 days after its last successful check.
An AI text studio on your own GPU
With an NVIDIA GPU of 6 GB or more, a local language model served by Ollama rewrites raw dictation into an email, a summary, a commit message or a prompt: 17 templates across 82 tones, plus templates you write yourself. Paste any text in and the same templates work on it.
Read-back and voiceover
On-device text-to-speech reads any text back in a natural voice, offline once the voice files are downloaded, and on Pro and Studio exports it as a WAV file. Read-aloud is not among the features OpenWhispr’s README lists.
Source: OpenWhispr README, OpenWhispr on GitHub. Checked on 27 October 2026.
An accuracy figure with its conditions
7.8% word error rate with Whisper large-v3 on an NVIDIA GeForce RTX 3080 Ti (12 GB), across 5 recorded clips, published on 16 July 2026 with the method and every clip, including the ones we lose on.
What it does not do: run on a Mac or Linux, transcribe meetings or audio and video files, or rewrite text with AI without an NVIDIA card, and its code is not open. The full comparison lists where other tools beat it.
OpenWhispr and Voxmelt, answered
Does OpenWhispr work on Windows?
Yes. OpenWhispr’s README lists a Windows installer (.exe) next to the macOS and Linux ones, and its latest release, 1.10.4, dated 21 October 2026 in its changelog, includes one. The README states no Windows version or hardware minimum; its local Whisper can use the GPU through CUDA, or through Vulkan on AMD and Intel. Voxmelt runs only on Windows 10 (64-bit, build 19041 or later) and Windows 11, and its default speech engine runs on the CPU.
Sources: OpenWhispr README, OpenWhispr on GitHub; Release 1.10.4, OpenWhispr on GitHub; Changelog, OpenWhispr on GitHub. Checked on 27 October 2026.
Is OpenWhispr free?
The app is. OpenWhispr calls itself the open-source and free alternative to WisprFlow and Granola, and its code is under the MIT licence, free for personal and commercial use. Its hosted service, OpenWhispr Cloud, is the default after onboarding; its changelog calls the usage meter the weekly OpenWhispr Cloud allowance and mentions paid plans and subscriptions. Its network notes also describe signed-out, fully local setups. Voxmelt starts with a 15-day trial of everything and then a free tier of 60 minutes of dictation a day, with no card; Pro is $49 once and Studio $99 once.
Sources: OpenWhispr README, OpenWhispr on GitHub; MIT License, OpenWhispr on GitHub; Network Allowlist, OpenWhispr on GitHub; Changelog, OpenWhispr on GitHub. Checked on 27 October 2026.
Does OpenWhispr send my voice to the cloud?
It depends on the models you pick. Its README offers fully private offline transcription with local models such as Whisper and NVIDIA Parakeet, where your audio never leaves your device, or cloud processing for speed. Its network notes list OpenWhispr Cloud as the default after onboarding: transcription, sync, agent reasoning, settings and usage go through its cloud API, and streaming transcription goes through one of four providers, Deepgram, OpenAI Realtime, AssemblyAI or Gemini Live. With your own API key, the work goes to that provider instead. Voxmelt has no cloud mode: your audio and transcripts never leave the PC.
Sources: OpenWhispr README, OpenWhispr on GitHub; Network Allowlist, OpenWhispr on GitHub. Checked on 27 October 2026.
Does OpenWhispr work offline?
With local models, yes: its README calls that fully private offline transcription. Its network notes say each local model downloads when you opt into it, so that first step needs the internet, and cloud processing needs a connection every time. Voxmelt needs the internet at setup, to download its speech model, and to download the AI model you pick if you use AI rewriting; after that it dictates with the Wi-Fi off, its AI rewriting runs on your own NVIDIA GPU, and a paid licence keeps working offline for 30 days after its last successful check.
Sources: OpenWhispr README, OpenWhispr on GitHub; Network Allowlist, OpenWhispr on GitHub. Checked on 27 October 2026.
Why pay for Voxmelt when OpenWhispr is free?
You may not need to. If OpenWhispr, set to local models, does what you need, it costs nothing. Voxmelt is for people who want local only by design, with no cloud default to change, plus an AI text studio of 17 templates across 82 tones on their own GPU, read-back with WAV export (Pro), and help by email at help@voxmelt.com. Try it before you decide: a 15-day trial of everything, then a free tier of 60 minutes of dictation a day, with no card, and $49 once if you keep it.
Is Voxmelt open source, like OpenWhispr?
No. OpenWhispr is open source under the MIT licence. Voxmelt is not: you get a licence to use it under our end-user licence agreement. It stands on open models, NVIDIA Parakeet, Whisper, Kokoro and the Ollama models you choose, which keep their own licences. If open source is a requirement, OpenWhispr is the pick.
Source: MIT License, OpenWhispr on GitHub. Checked on 27 October 2026.
Which is more accurate, OpenWhispr or Voxmelt?
There is no like-for-like answer. Both can run the same open speech models: OpenWhispr’s README names Whisper and NVIDIA Parakeet among its local models, and Voxmelt dictates with NVIDIA Parakeet v3 on the CPU, with Whisper as the opt-in engine. Voxmelt publishes a 7.8% word error rate with Whisper large-v3 on an NVIDIA GeForce RTX 3080 Ti (12 GB), across 5 recorded clips, dated 16 July 2026; the clips cover clean dictation, technical / code speech, fast natural speech with filler, medical jargon and background noise, and the method is on our benchmark page. Our benchmark did not test OpenWhispr, and its README publishes no accuracy figure. The fairest test is your own voice: dictate the same paragraph into each and count the fixes.
Source: OpenWhispr README, OpenWhispr on GitHub. Checked on 27 October 2026.
Try it next to OpenWhispr
Install both on the same PC, set OpenWhispr to a local model, dictate the same paragraph into each, then turn the Wi-Fi off and try again. Free tier with no card, a 15-day trial of everything, and $49 once after that.
Direct installer, for PCs without the Store; Windows will warn because it is not code-signed.
Read how offline dictation works on Windows and the free routes to private dictation on Windows 11, or see Voxmelt against Wispr Flow, Superwhisper, Otter AI and Dragon.
Where every OpenWhispr fact comes from
OpenWhispr’s own GitHub repository: its README, its network notes, its changelog, its licence and its latest release, read on the date given. If one of them has changed since, tell us at help@voxmelt.com and we will fix this page.
- OpenWhispr README, OpenWhispr on GitHub. Checked on 27 October 2026.
- Network Allowlist, OpenWhispr on GitHub. Checked on 27 October 2026.
- Changelog, OpenWhispr on GitHub. Checked on 27 October 2026.
- Release 1.10.4, OpenWhispr on GitHub. Checked on 27 October 2026.
- MIT License, OpenWhispr on GitHub. Checked on 27 October 2026.
Voxmelt’s figures come from our benchmark, published 16 July 2026, and our pricing page, and what can leave your PC is listed in our privacy policy.