SnailText

SnailText is voice-to-text that runs entirely on your own machine. Press a hotkey, speak, and clean text lands wherever your cursor is - in any app, on Mac or Windows. Nothing you say is uploaded. No account needed to start. Most dictation tools stream your audio to a server to transcribe it. That's fine for a grocery list, less fine for source code, client names, an NDA'd spec, or a performance review. SnailText was built the other way around: the speech model (OpenAI Whisper or NVIDIA Parakeet) runs locally, your audio is processed in RAM and never written to disk, and the moment the text is ready the audio is gone. Privacy by architecture, not by policy - you can verify it in your own network monitor, because there's nothing to see. WHAT MAKES IT DIFFERENT 100% on-device. Works on a plane, in a tunnel, on a locked-down work laptop. No internet required after the model downloads. Mac AND Windows, as equal first-class citizens. Most local dictation apps are Mac-only. GPU-accelerated on both (Metal on Mac, Vulkan on Windows), with a CPU fallback that's still real-time on compact models. It's genuinely AI dictation, locally. A second local model (Gemma) cleans up the raw transcript: it drops filler words ("um", "you know"), fixes punctuation and capitalization, and restores code identifiers in the casing you want (say "recording completed", get recording_completed or recordingCompleted). Competitors send this cleanup step to GPT/Claude/Gemini in the cloud. SnailText keeps it on your machine. Works in every text field. Slack, Cursor, VS Code, Gmail, Notion, your terminal, a browser form - if you can type there, you can dictate there. No per-app plugin. A real free tier, not a trial. Unlimited dictation on the two compact Whisper models, no word cap, no account, no card. Pro unlocks larger models and the on-device AI cleanup. NEW: BATCH FILE TRANSCRIPTION (rolling out soon) You already have a folder of call recordings, voice memos, or interview audio sitting on your disk. Point SnailText at a single file or a whole folder and it transcribes the lot - locally, no upload, on the same model you already use for dictation. Choose your output: one transcript per file, or a single combined document, as .txt, .md, or .json. Optionally run each transcript through the on-device AI cleanup. Audio formats ship first (mp3, m4a, wav, ogg, flac); video (mp4, mkv, mov, webm) follows. It's the "batch of recordings to text" workflow you'd normally hand to a cloud service like MacWhisper or a paid transcription site - except it runs on your own hardware and nothing leaves the building. BUILT FOR PEOPLE WHO DICTATE A LOT Developers and vibe-coders: dictate prompts to Cursor/Copilot, commit messages, code comments, PR descriptions. The Development profile keeps snake_case and camelCase intact. Writers, students, and researchers: draft at the speed of speech, then let the cleanup pass make it readable. Anyone with RSI or who just types too much: your hands get a break without your workflow breaking. Privacy-conscious professionals (legal, medical-adjacent, finance): dictation that's compliant by design because the data never leaves your device. No BAA needed when there's no server. MORE UNDER THE HOOD Custom dictionary + text snippets so your product names, jargon, and boilerplate come out right. Voice-activity detection and a multi-layer hallucination filter, so silence and background noise don't turn into phantom text. Streaming transcription: on a GPU, long dictations are ready a second or two after you stop talking, because inference happens while you speak. Searchable local history (stored in SQLite on your machine, deletable and exportable on your terms) with usage stats. Interface in 7 languages (EN, RU, ES, PT, FR, DE, NL); dictation in ~99 languages via Whisper. Ergonomic hotkeys: double-tap a modifier, a mouse side button, or push-to-talk - whatever fits your hands. PRICING Free to start, no account needed. Pro is $7.49/month or $8