My own dictated messages from building it, kept as I said them. Facts in the boxes.
The 1-minute version
- Voice typing on a Mac falls apart when you speak Hinglish, half Hindi and half English. So I built my own, Vaani. It's free and runs on the Mac itself.
- I test it on of my own real dictations, replayed every time a rule changes.
- Reading short clips a second time cut the wrong words from .
- Two "smart" extras together doubled the wait, , and one of them made mistakes worse. That one is off now.
- It still mishears its own name sometimes.
Why I needed my own
For a few days, I've been using Vaani and it is not getting the exact words that I am saying. I don't know what the issue is. Maybe the model that I'm using is outdated.
And this thing does not happen when I use the default voice typing in Claude, or any other tool that I used earlier. So why is it happening in the tool that we have just built?
It can't spell its own name
Exhibit A
for a few days, I've been using Vani and it is not getting exact words
23 Sepauto applies those fixes properly into the Wani app
28 SepI would want to get money ready for the website
1 Octthe waitlist that you built for Wani for Windows
22 Oct
So I pasted my recent messages, the ones I had fixed by hand, next to what the app typed. The fixed ones are what I really meant, and the transcription is the wrong one that it takes. So when you match it, you can find what I really meant and what it took from me.
Another bug I found: I am saying Hinglish and it is typing English.
Where it got stuck
The bubble stays frozen on the screen. Sometimes the recording is done, but when I press the dictate button again, it keeps on loading and nothing is pasted anywhere. And pressing the key sometimes does nothing. So these are some of the bugs.
And the preview: when I am speaking and the text is being previewed, the animation was utter nonsense. It was so stagnant, so jittery, not smooth. I am speaking and the words are being typed there, so it should be like a magical thing that is happening.
Tested on my own voice
We are not using any AI for corrections like many big apps do, because we need more speed. So for speed, we have created this file of corrections that we keep updating.
6.1%words wrong
down from 8.6%, after a second read of short clips
When you stop talking, a clip of 15 seconds or less gets read again as a whole. It costs 0.6 seconds on a 2-second clip and 5.4 seconds on a 54-second one, so long clips skip it. Measured on my 6 real test clips, 23 Sep 2026, so a small test.
The "smart" extras, raced
Wait after I stop talking, averaged over 7 of my clips.
A replay of my measured times, 24 Sep 2026. Sample line; no microphone, nothing recorded.
The word list was the slow one, and it also made mistakes go up on my normal speech. It's off by default now. The second read costs almost nothing, so it stays on.
What it does now
And converting Vaani into a text cleaner and text formatter: bullets, numbering, all the possible things that can be done with text through Vaani.
What I do now
- Test on
- My own voice: 536 real dictations, not clean sample audio.
- Keep a feature
- Only when the numbers say so. A smart-sounding one got switched off.
- Feels worse?
- Check the settings before the code.
Sources
- HiACC Hinglish speech corpus paper (2025)
- Hindi-English Code-Switching Speech Corpus (2018)
- Interspeech 2021, Hindi in Latin letters
- My own tests, 23 and 24 Sep 2026
How this post was made: the main text is my own dictated messages from building Vaani, kept close to how I said them. Claude wrote the headings and the boxes from my test notes and the linked sources, Gemini checked the facts, and I read and approved it. Numbers are from my own tests in September 2026. This version of my words was remixed by AI; the numbers and boxes are unchanged.