Offline dictation on Windows with Parakeet v3 int8 on CPU (my Handy fork) - plus a batch tab for files
Dev here, and full disclosure up front: this is my own paid app, and it started as a fork of Handy (MIT). Posting it because the CPU-only part is the bit this sub actually cares about, and because Handy deserves credit for the base. The problem I had: every…
Dev here, and full disclosure up front: this is my own paid app, and it started as a fork of Handy (MIT). Posting it because the CPU-only part is the bit this sub actually cares about, and because Handy deserves credit for the base. The problem I had: every dictation tool I liked was either a subscription with cloud STT, or wanted a GPU sitting there hot just so I can talk into a text field. I type for a living, my wrists complain, and I didn't want my voice leaving the machine. What ended up working: Parakeet TDT 0.6B v3, int8 ONNX (~478 MB, pulled on first run). CPU-first, no CUDA path at all in v1 - it runs fine on an old laptop. Silero VAD for endpointing (straight from upstream Handy). Push-to-talk hotkey, transcribe, paste into whatever window has focus. Post-processing is plain rules, not an LLM: three presets (prose / code / email) for punctuation, capitals and filler removal, plus a personal dictionary for hotwords and a snippet expander. The part I use more than I expected is a batch tab - drop a folder of audio or video (wav/mp3/m4a/flac/ogg/opus, mp4/mov/mkv/webm) and it writes a .txt next to each file with the same model. Old voice memos and meeting recordings went from "someday" to done in one evening. Honest limits, since you'd find them anyway: Windows 10/11 x64 only. Tauri would allow Mac/Linux, I just haven't done it. Rules-based cleanup won't repair a mangled sentence the way an LLM pass would. Optional local polish is a v1.1 idea, not shipped. 25 languages from the model (EU set plus ru/uk), not the "100+" a cloud service claims. It's paid: free tier is 21 dictations a day with no time limit, licence is one-time, no subscription. Link if you want to poke at it: https://www.softorbits.net/voice-dictation-software/ Mostly I'm curious what people here actually run for local dictation day to day. Is anyone still on whisper.cpp for this, or has the Parakeet family taken that slot for you too? And if you've tried a small local LLM for transcript cleanup instead of rules, where did it genuinely earn its keep?
已收录讨论
LLM-generated posts are bad.