Original Summary

I built a macOS dictation tool over the last few months because every one I tried mangled Indian names. Hold right Option, talk, release, and cleaned up text lands at the cursor in whatever app you are in. Whisper transcribes locally, then an LLM pass fixes it up. I am at the point where it works well enough that I use it all day, and I genuinely cannot tell whether that means anything to anyone else. So: numbers first, then the full feature list, then the questions I actually want answered. WHERE IT STARTED AND WHERE IT IS NOW Measured on my own benchmark, which I will describe honestly so you can discount it: five sentences synthesized with two macOS voices, one standard and one Indian English, run through the real pipeline, word error rate against the reference text. Raw transcription came out at 4.1 percent word error rate. After the cleanup stage it is 2.4 percent, so the cleanup is removing roughly 40 percent of the remaining errors rather than just tidying punctuation. Proper nouns were the real problem. Names, Indian cities, company names. With the vocabulary pack and the learned dictionary, accented proper nouns went from 85 percent correct to 100 percent on my name benchmark. That is the number I actually care about, because a transcript with the right words and the wrong name is still unusable. Speed: about 1.5 seconds to transcribe plus 1.5 seconds to clean. Long notes still come back in 2 to 3 seconds from release, because both stages stream while you are still talking rather than starting when you stop. Cost per dictation dropped about 70 percent once I made the system prompt byte identical between calls so the provider caches it. Roughly 0.006 dollars a minute of audio now. A heavy day is under a dollar. There are 42 tests, and every regression that nearly shipped is pinned as a case, mostly around streaming seams and the clipboard restore. What I have not done: real human speech on a standard corpus. That is the honest test and my benchmark is not it. THE FEATURE LIST Hold a global hotkey to talk, or quick tap to lock hands free, or just stop talking and a 2.5 second pause ends the recording. Cleanup removes fillers and false starts, fixes punctuation, and resolves spoken self corrections, so "send it to John, no wait, Sarah" comes out as one sentence about Sarah. Spoken commands like "new line" are applied rather than typed out. Tone matches the app you are typing into, so a Slack message and a document do not come out identical. Snippets. A spoken trigger expands to saved text. A personal dictionary that both biases the recognizer and tells the cleanup model the exact spellings. It learns without being configured. On first run it reads the Mac's own contact card, so your employer name and work email domain are spelled right from the first dictation instead of guessed phonetically. Then names from Contacts, proper nouns visible on screen while you dictate, a regional vocabulary pack from your locale, and a

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/26 15:26:38