What does voice input for the terminal look like in practice?
Voice now carries a lot of my terminal input. Long instructions to a coding agent are slow to type and easy to say out loud, so I dictate them into the terminal and keep the keyboard for the rest.
The setup has three parts. A dictation app turns speech into text, and I use two: Wispr Flow and DictationPal. A terminal receives it, which for me is Ghostty (its pane layout is another setup in this series). A counter tells me, per day, how many characters came from voice and how many from the keyboard. The counter turned out to matter most, because without it I had an impression (“I dictate a lot”) and nothing to check the impression against.
How do I measure the voice share?
I count on one Mac only. The terminal is Ghostty. For voice I use completed transcriptions from Wispr Flow and from DictationPal, and I count the characters in each one that is associated with Ghostty. For typing I count characters entered on the hardware keyboard. Pasted text is excluded, because pasting a log file into a prompt says nothing about how I produced input.
The voice share for a day is voice characters divided by voice plus typed characters. On October 4 that was 40,161 voice characters out of 72,618 counted, which is 55.3%. The remaining 32,457 were typed.
That was the highest complete day in the record so far. The typing data starts on September 21, so the useful window is short. Across the 14 complete days from September 21 to October 4, voice made up 122,821 of 515,393 characters, or 23.8%. Single days ranged from 1.0% to 55.3%. October 5 shows 69.6%, but it was a partial day when I pulled the data, so I do not use it.
The spread is the main finding. My voice share is not a stable habit yet. The data shows that it swings from day to day, and it does not tell me why.
What does this count get wrong?
Three things.
First, the count is of transcriptions, not of text confirmed to land in the terminal. A completed DictationPal record does not prove that the text was pasted anywhere. The voice figure is therefore more likely too high than too low.
Second, typing coverage starts on September 21. The 90-day chart on the Compounding CEO page has voice data from July 8, but the earlier days have no typed counterpart, so no share can be computed for them.
Third, this is characters, which is input volume. It does not measure time saved, and it does not measure whether the prompts were better. A weekly loop and an email report now run on the numbers, so over time I can compare weeks, but a comparison of shares will still not answer the speed question.
What does it cost?
Measurement effort, mostly. Wispr Flow and DictationPal are separate apps with separate histories, which is why I count both and why a single day’s number needs the caveats above. Misheard words add a second cost. Another setup page in this series covers how I handle one recurring misrecognition: the spoken command to press Enter in Wispr Flow.
Dictation also needs a place where you can talk out loud. I have not measured that cost, and it limits where the setup works for anyone.
Who is this for?
It fits anyone who writes long instructions to AI agents and can talk out loud while working. If most of your terminal time goes to short commands, expect the share to stay low.
It also suits people who like to check their own habits against a number. The measurement is the part that transfers. Even if you never dictate a word, counting where your input comes from for two weeks shows how much of your terminal time goes to composing text.
How can I start in 15 minutes?
Pick one dictation app and install it. Use it only for the next five prompts you would normally type to a coding agent, and leave short commands on the keyboard. Notice where it misses words and note them down.
Then decide what to count. The simplest version is a tally at the end of the day: roughly how many prompts were spoken and how many typed. A character count is better, but it needs a logger on both sides, and I would not build that before you know you will keep dictating. Write down the date you start, because a share without a coverage date is hard to read later.
After two weeks, compare the days. If the share swings as widely as mine did, look at what the high days had in common before you try to lift the low ones.
Where does this fit in the series?
This is one practice from the Compounding CEO playbook, where the voice-input chart sits with the other measured entries. The practice card is Let voice carry more of the input.
Frequently Asked Questions
How much of my terminal input can voice realistically cover?
On my Mac the daily share of counted characters ran from 1.0% to 55.3% between September 21 and October 4, and the average over those 14 complete days was 23.8%. The share depends heavily on the kind of work that day. Long prompts to coding agents are easy to dictate, while short commands and edits stay on the keyboard.
Which tools do you count for voice input?
I count completed transcriptions from Wispr Flow and DictationPal that are associated with the Ghostty terminal on one Mac. Typed characters come from the hardware keyboard, and pasted text is excluded. Other apps and other devices are outside the count.
Does a higher voice share mean I work faster?
No. The figure measures how input was produced, in characters, and says nothing about time saved or output quality. Speed and quality are separate questions that need their own measures.
Can the dictation count be wrong?
Yes. A completed transcription is counted even though the app does not confirm the text landed in the terminal. A DictationPal record in particular does not prove the text was pasted, so the voice share is probably a little high.
Ready to Transform Your AI Strategy?
Get personalized guidance from someone who's led AI initiatives at Adidas, Sweetgreen, and 50+ Fortune 500 projects.
