Tested
Dictation with an Indian English accent: four engines, one voice
I speak Indian English. Before choosing a default speech engine for Precis, I read the same 14 hard phrases into four engines and counted the mistakes. Here's what happened, with the limits stated.
Short answer: On my voice, NVIDIA Parakeet made the fewest mistakes (6.7% of words wrong), then Whisper (8.5%), Apple's speech model (9.7%) and Apple's dictation model (10.3% set to US English, 17% set to English (India)). One speaker and 14 phrases is a data point, not a benchmark.
| Engine | Words wrong | Notes |
|---|---|---|
| NVIDIA Parakeet (Precis's default for English) | 6.7% | Runs on your Mac. English and 24 European languages. |
| OpenAI Whisper (in Precis) | 8.5% | Runs on your Mac. 99 languages. |
| Apple's speech model (SpeechTranscriber) | 9.7% | The same score set to US or Indian English. |
| Apple's dictation model (DictationTranscriber) | 10.3% set to US English, 17% set to Indian English | Apple says it uses the same models as system dictation. |
My own test, run on 2 October 2026 with Precis's engine check. Apple describes its dictation model in the DictationTranscriber documentation.
How I tested
I read 14 phrases from a list written to be hard: names (Nikesh, Priya, Karthik, Rahul), times and prices ("5:30 pm on Friday", "29 dollars"), a self-correction ("Tuesday, no wait, Wednesday"), fillers ("um, so basically I think we should, uh, launch it") and tech words (Supabase, Vercel, SwiftUI, GitHub). Every engine heard the same recordings. Each recording was matched to the phrase it was closest to and scored by the share of words it got wrong. I ran it on my M4 Mac on 2 October 2026.
What every engine missed
"What's the difference between let and var in Swift" and "keep it under 300 words" were missed by all four. Switching engines wouldn't have fixed them: that's my accent and microphone. It's also what a dictionary is for. Add the words an engine keeps missing, and they're spelled your way from then on.
Two surprises
- "English (India)" made Apple's dictation model worse on my voice: 17% of words wrong, against 10.3% set to US English. Apple's speech model scored the same either way. Test both settings with your own voice before assuming the local one helps.
- Boosting tech names inside the engine made it worse. Telling Parakeet to listen for names turned "Swift" into "SwiftUI" and "Python script" into "Python TypeScript". So Precis fixes names in the text afterwards instead: ChatGPT, iOS, macOS, Supabase, GitHub and SwiftUI are written correctly, with your own Dictionary on top.
What I'd do with an Indian accent
- Try an app with Parakeet first. In Precis it's the default for English.
- Test Apple Dictation with both US English and English (India) before picking one.
- Add the words it keeps missing to the dictionary, especially names.
Limits
- One speaker, one microphone, one quiet room and 14 phrases. Your voice will give different numbers.
- I didn't test cloud services such as Wispr Flow, so this says nothing about them.
- Hindi, Tamil and other Indian languages weren't part of this test. Whisper can transcribe them, but Apple's model doesn't support them for Styles, so Precis types them cleaned up without rewriting.
Questions
Which dictation engine works best for an Indian English accent on Mac?
On my voice, NVIDIA Parakeet got the fewest words wrong (6.7%), then Whisper (8.5%), Apple's speech model (9.7%) and Apple's dictation model (10.3% set to US English). That's one speaker and 14 phrases, so test your own voice.
Does setting English (India) improve Apple's dictation?
Not on my voice: Apple's dictation model got 17% of words wrong set to English (India), against 10.3% set to US English. Apple's speech model scored the same either way.
Can Precis type Hindi or Tamil?
Whisper in Precis can transcribe them. Apple's model doesn't support them for Styles, so the text is cleaned up but not rewritten.
Try Precis free for 7 days
Hold ⌥ Space in any app and talk. No card and no account, then $39 once for 2 Macs.
Free for 7 days, no card. Needs macOS 26.5 on Apple silicon.