Croatian voice dictation intermittently transcribes in Cyrillic instead of Latin script

What issue are you seeing?

Voice dictation for Croatian intermittently produces Cyrillic text instead of Croatian Latin script.

This happens when using the built-in microphone/dictation control in OpenAI clients. I have observed it on both:

• Codex / ChatGPT on Windows desktop
• ChatGPT on Android

The same user can dictate Croatian and sometimes receive normal Latin-script Croatian, while another dictation is rendered in Cyrillic. There is no intentional language or script switching.

This is particularly disruptive because standard Croatian uses the Latin script, and Cyrillic output can be difficult or effectively unreadable for a Croatian user.

Useful control test

On the same Android phone, if I use Android/system keyboard voice typing instead of the microphone built into ChatGPT, Croatian is transcribed normally in Latin script.

So the behavior appears specific to the OpenAI in-app dictation/transcription path rather than the phone microphone or Croatian speech itself.

Steps / observed sequence

1. Open ChatGPT or Codex and use the application's own microphone/dictation control.
2. Speak normally in Croatian.
3. Repeat several dictations in Croatian.
4. Some transcripts are correctly returned in Latin script.
5. Intermittently, a transcript is instead returned in Cyrillic.
6. On Android, repeat the same type of Croatian dictation through the system keyboard's microphone; it returns Latin script normally.

The issue is intermittent rather than a deterministic one-utterance reproducer.

Expected behavior

Croatian speech should consistently be transcribed as Croatian using Latin script unless the user explicitly requests another language/script.

An established Croatian-language conversation should also provide a strong signal against switching the transcript to Cyrillic.

Suggested improvement / fallback

Please consider exposing an explicit dictation language/script preference, for example:

• Croatian (Latin script)
• or a broader Always use Latin script transcription preference

This would both prevent incorrect script auto-detection and give users a reliable fallback when automatic language/script detection is ambiguous.

Related issue

This appears conceptually related to #44926, where Czech voice input in the Codex Windows desktop app intermittently switches to unrelated languages/scripts. This report is specifically about Croatian repeatedly switching between the expected Latin script and Cyrillic, and it also reproduces in ChatGPT on Android.

I can provide examples/screenshots and exact app versions from affected devices if useful.

No repository file, test, or entry point is identified. First reproduce repeated Croatian dictation in Codex or ChatGPT and compare it with Android system keyboard voice typing, noting when Cyrillic appears. Done means Croatian dictation consistently uses Latin script or provides the requested language/script preference.