How I used Codex to investigate and fix the issue where Superwhisper's Voice to Text keeps reverting to English/Ultra

When writing documents on a Mac, writing blog posts, or consulting with ChatGPT or Codex, inputting by voice is overwhelmingly faster than typing by hand.

For that reason, I use Superwhisper quite a lot.

However, there was something that had been bothering me for a while.

Even though I want to use Superwhisper in Japanese Standard, I notice that the settings screen reverts to English or Ultra.

Every time before I use it, I have to change the Language back to Japanese and the Voice Model back to Standard.

This is subtly exhausting.

Voice input is valuable because you want to use it the moment an idea strikes. Yet, having to open the settings screen to fix it every time is quite stressful.

This time, I consulted Codex in natural language to investigate this issue, and I'm writing about how I managed to get it to a very practical state.

What was happening

The symptoms were roughly as follows.

• When launching Superwhisper, the Language reverts to English

• The Voice Model sometimes reverts to Ultra

• Even though I want to use Japanese Standard, I have to fix the settings every time

• There is no apparent 'Save Settings' button

• Recording history had also accumulated to several gigabytes

At first, I honestly suspected, 'Are they making it inconvenient on purpose to push free users toward the paid version?'

But when I actually looked into it, the cause was a bit more complicated.

In conclusion, Superwhisper's Default mode was behaving as if it were regenerating with English + Ultra settings upon startup.

Moreover, even if I corrected the configuration file to Japanese Standard, the Default would revert to English Ultra settings again when the app was restarted.

This was the core of the problem.

The configuration file found first

When I had Codex search my Mac, it found the configuration folder related to Superwhisper.

The main location is here.

/Users/Claw/Documents/superwhisper

Inside it, there were configuration files for each mode.

/Users/Claw/Documents/superwhisper/modes/default.json
/Users/Claw/Documents/superwhisper/modes/voice to text.json

The first problem I found was that the `voice to text.json` file was set to something like English + Ultra Cloud.

So, I first corrected this mode for Japanese Standard.

Specifically, roughly speaking, these are the settings.

{
  "language": "ja",
  "voiceModelID": "small",
  "languageModelID": "",
  "translateToEnglish": false
}

At this point, I thought, "This might have fixed it."

However, the story continued a bit further.

The real cause was the regeneration of the Default mode

The next day, when I launched Superwhisper, it had reverted to English again.

Moreover, on the screen, it looked like the Default mode was selected, with the Language set to English and the Voice Model set back to Ultra or Standard.

So, I had Codex investigate it again.

And this is what I found out.

`voice to text.json` is fixed.

But when I launch Superwhisper, `default.json` is automatically regenerated.

The contents of the regenerated Default were in this state.

{
  "key": "default",
  "name": "Default",
  "language": "en",
  "voiceModelID": "sw-ultra-cloud-v1-east"
}

In other words, the true identity of the English + Ultra I was seeing every time was this Default mode.

Even if I corrected it to Japanese / Standard in the settings screen every time, the Default would reappear as English Ultra-based upon startup.

No wonder it kept reverting every time.

What I asked Codex to do

I didn't ask Codex to fix any difficult code.

I just explained it in quite natural Japanese, something like this.

Superwhisperを使っていると、毎回設定がEnglishとUltraに戻ってしまいます。
日本語のStandardで使いたいのに、毎回直すのが面倒です。
Macの中に設定ファイルがあるなら、原因を調べてください。
バックアップを取ったうえで、安全に直せるなら直してください。

I didn't know where the configuration files were from the start.

I didn't understand the contents of `default.json` or `voice to text.json` either.

I simply asked, "I'm having trouble with these symptoms. Please find the cause and fix it safely."

From there, Codex handled it by:

• Searching for Superwhisper-related files

• Finding the configuration folder

• Checking the mode configuration file

• Identifying the problematic setting

• Taking a backup instead of overwriting immediately

• Rewriting the settings for Japanese Standard

• Explaining what was changed after the fix

It followed this process to resolve the issue.

This was extremely convenient.

Rather than "writing code with AI," it felt more like consulting in natural language to fix Mac settings that I was afraid to touch myself.

The final solution implemented

The final solution I implemented is quite practical.

First, I rewrote `default.json` for Japanese Standard.

{
  "key": "default",
  "name": "日本語 Standard 無料",
  "language": "ja",
  "voiceModelID": "small",
  "languageModelID": "",
  "translateToEnglish": false
}

However, this alone might still result in it being overwritten upon startup.

In fact, Superwhisper was regenerating the Default settings when it launched.

So, while consulting with Codex, I used the macOS file locking feature to make it difficult for the app to automatically overwrite 'default.json'.

Specifically, I applied the 'uchg' lock attribute.

uchg /Users/Claw/Documents/superwhisper/modes/default.json

With this, even if Superwhisper tries to regenerate the Default settings to English Ultra at startup, it generally cannot overwrite them.

Furthermore, as an additional safeguard, I also created a fixed Mode.

日本語 Standard 無料 固定

This Mode is also set to Japanese Standard.

{
  "language": "ja",
  "voiceModelID": "small",
  "languageModelID": "",
  "translateToEnglish": false
}

In other words, in the end, I implemented the following measures:

• Set Default to Japanese Standard as well

• Lock the Default settings to prevent them from reverting to English Ultra

• Create a fixed Japanese Standard Mode as a backup

That is the approach I took.

The screen display may still appear as English

I will be honest about this part.

While the configuration file is now set to Japanese Standard,

checking the recording history confirms that the actual processing is running in Standard mode.

However, on the Superwhisper screen, the Language sometimes appears as English.

This is a bit unsettling.

But when I actually perform voice input, it is recognized in Japanese.

In short, this is the current state of things.

• It may appear as English on the screen display

• But it is Japanese in the configuration file

• The actual Voice Model is Standard

• Actual voice input also works in Japanese

• The issue of reverting to Ultra has been significantly suppressed

Therefore, it is more accurate to say that I brought it to a state where it can be used with Japanese Standard for practical purposes, rather than saying the UI display was perfectly fixed.

Exaggerating this would lead to a loss of credibility, so it is better to write that I "identified the cause and addressed it to a state where it is not a practical problem" rather than a "complete solution."

There was also a storage issue with recording history

Another thing I was struggling with was the storage capacity of the recording history.

I had been using Superwhisper quite a lot, so about 6GB of recording history had accumulated.

The storage location was here.

/Users/Claw/Documents/superwhisper/recordings

I rarely look back at my past recording history.

For voice input, it is enough if it becomes text on the spot.

Therefore, there is not much need to keep old recording history.

However, I am afraid to delete it myself from Finder.

If I accidentally delete configuration files or Mode settings, the settings I worked hard to fix might break.

So, I asked Codex for help with this as well.

Superwhisperの録音履歴が6GBくらい溜まっています。
私は履歴を見返さないので、古い録音は削除したいです。
ただし、設定ファイルやMode設定は消したくありません。
録音履歴だけを安全に削除する方法を考えてください。

As a result of Codex checking it, the only thing that seemed safe to delete was the recording history folder.

/Users/Claw/Documents/superwhisper/recordings

Conversely, I was also able to organize which locations should not be touched.

/Users/Claw/Documents/superwhisper/settings
/Users/Claw/Documents/superwhisper/modes
/Users/Claw/Documents/superwhisper/models
/Users/Claw/Library/Application Support/superwhisper

It is scary to make these kinds of judgments on my own, but if I tell Codex my goal, such as "I don't want to delete the settings, I only want to delete the recording history," I can organize it quite safely.

You can also perform regular cleanup with Codex automation

Recording history may continue to grow in the future.

So, I also created a plan to clean it up periodically using Codex's automation feature.

The details are as follows.

• Execute every Sunday at 3:00 AM

• Target only the Superwhisper recording history folder

• Delete only recording folders older than 30 days

• Do not touch settings, modes, models, or Application Support

Even though it runs every week, it doesn't mean everything is deleted every week.

The deletion target is strictly limited to recordings older than 30 days.

It is designed to automatically clean up only recordings that are too old while keeping recent history.

I didn't write these difficult settings myself either.

I just asked like this.

Superwhisperの録音履歴が今後も増え続けると思います。
毎回手動で消すのは面倒です。
設定ファイルには触らず、録音履歴だけを対象にして、
30日より古い録音だけを定期的に削除するようにできますか？

In response, Codex organized which folders to target, how old the items to be deleted should be, and how frequently to run it, and proposed it as an automation.

Being able to request this kind of mundane maintenance in natural language is quite practical.

What I learned this time

The biggest takeaway from this experience was not just "having AI write text," but the feeling of "being able to investigate and fix troublesome issues happening on my Mac together."

Of course, you can't completely leave everything to it.

This time as well, my initial assessment was that `voice to text.json` was the cause.

However, after it recurred and I investigated further, I discovered the behavior where the default is regenerated as English Ultra upon startup.

In other words, Codex didn't get it perfectly right on the first try.

However, this is actually what makes it feel like real-world work.

In reality, troubleshooting often doesn't end with a single attempt.

Formulate a hypothesis.

Verify it.

If it's wrong, correct it.

Look one level deeper.

Finally, bring it to a solution that works in practice.

It was convenient to be able to proceed through this flow while consulting in natural language.

Summary

To sum up what I did this time, it's a story about how I dealt with Superwhisper's configuration issues and recording history capacity problems by consulting with Codex in natural language.

Here is what I ultimately found out:

• Superwhisper settings are saved per Mode

• English Ultra-related settings remained in `voice to text.json`

• Furthermore, the core issue was that `default.json` was being regenerated as English Ultra at startup

• I changed the Default to Japanese Standard and locked it to prevent overwriting

• Although it may appear as English on the screen, it is now in a state where the actual processing runs on Japanese Standard

• Recording history accumulates in `recordings`

• You can also create an automation that periodically deletes only the recording history without touching the configuration files

What I did was not to understand every single technical command.

"I'm having this kind of trouble."

"I want it to be like this."

"Please do it safely without touching anything dangerous."

I just told Codex this.

Then, Codex found the files, investigated the cause, made backups, fixed the settings, and even helped me organize my recording history and automate tasks.

When people think of AI tools, they often think of text generation or code generation.

But what I felt this time was the convenience of being able to fix computer troubles I didn't understand by discussing them in natural language.

AI voice input tools like Superwhisper are truly convenient.

However, when you use them for a long time, minor issues arise, such as settings reverting, strange modes remaining, or recording history piling up.

In those times, instead of just searching and worrying alone, you can consult Codex in natural language and have it help you with everything from investigating the cause to fixing it and automating the process.

I felt this was a very practical way to use it.

And as in this case, even if the initial assessment is slightly off, you can re-examine and dig deeper into the cause.

I think this is also a very important point for AI used in practical work.