Voxmelt Reviews and Details

Questions & Answers

1. Why should a person choose Voxmelt over its competitors?

   Because the whole loop runs on your machine. Wispr Flow, Superwhisper's cloud modes, and Otter process your voice on their servers; Windows Voice Typing sends audio to Microsoft; Dragon costs $200-700 up front. Voxmelt runs dictation on your CPU and AI cleanup on your GPU, so zero audio bytes leave your PC - and you can verify that claim in sixty seconds with any network monitor, which is not a test the cloud tools can pass. You also get the AI text studio and local voiceover in the same subscription, where most tools charge separately or don't offer them at all. Free tier is 60 min/day with no card.

2. How would you describe the primary audience of Voxmelt?

   People who dictate for a living and can't - or won't - send their voice to a cloud server. In practice that's four groups: developers and technical writers who talk through specs, commits, and tickets; clinicians and legal/finance professionals where the recording itself is regulated (HIPAA, privilege, data residency); privacy- conscious power users who already run local models via Ollama and want dictation to match; and air-gapped or field users - defence, research, secure labs - where there is no network to upload to. The common thread isn't a job title, it's a hard rule that sensitive audio stays on the machine. Since v1.18 removed the GPU requirement, that audience is now anyone with a Windows 10/11 laptop, not just people with an RTX card.

3. What's the story behind Voxmelt?

   For forty years we adapted to computers: click here, type there, learn the menus. That era is closing. In the cognitive intelligence era the machine meets you at language - you say what you mean, and the model does the translating. Talking becomes the keyboard. Intent becomes the interface.

   Everyone agrees on that much. The interesting question is where the intelligence sits. Streaming your voice to a datacenter is one answer, and it happens to suit the people who own the datacenters. We took the other one: the models are small enough now, the hardware on your desk is strong enough now, and the most personal interface ever built should run on the most personal computer you own.

   Voxmelt is built by one person, Param Nimbark, and stays private because of how it's built rather than because of a policy anyone wrote.

4. Which are the primary technologies used for building Voxmelt?

   Tauri v2 with a Rust backend and a React + TypeScript frontend, which is why the installer is ~12 MB instead of a few hundred - no Electron. Speech recognition runs through Python sidecars: NVIDIA Parakeet TDT 0.6B v3 as int8 ONNX via onnx-asr on the CPU (the default), and OpenAI's Whisper via faster-whisper / CTranslate2 on CPU or CUDA for the ~100-language cases. AI text processing calls a local Ollama model over localhost - bring any of Gemma, Llama, Qwen, or Mistral. Text-to-speech is Kokoro, also local. Accounts and licensing are Supabase; billing is Razorpay. Every model runs on the user's machine; the only network calls are model download, licensing, and update checks.

5. What makes Voxmelt unique?

   Voxmelt is the only Windows dictation app that splits the work across your own hardware: NVIDIA's Parakeet v3 transcribes on your CPU, so dictation needs no graphics card at all, and if you do have an NVIDIA GPU it stays free to run a local LLM that rewrites, summarizes, or translates the text in one click. Most tools do speech-to-text and stop, or do the AI part in the cloud. Voxmelt does both on hardware you own: audio and text never leave the machine, there's no per-minute meter, and it keeps working with the network cable unplugged. It also reads text back in 32 local voices with WAV export.

6. Who are some of the biggest customers of Voxmelt?

   Voxmelt is early - publicly launched July 2026.