Summary:
There is an ongoing, systemic issue with the speech-to-text interface in the ChatGPT mobile app, specifically when using the built-in microphone button (i.e., not relying on third-party keyboards or tools). The app consistently misidentifies spoken English as Welsh. This occurs despite every visible language setting being configured for English and all automatic language detection being disabled wherever possible.
Details:
The device in use is entirely configured for English (UK), with no Welsh language preferences, inputs, or history.
The app misidentifies input as Welsh only within ChatGPT—this does not occur with other apps or keyboard-based dictation tools.
The issue persists across updates and across months of use, indicating a deeper failure in the speech recognition framework used by the app itself.
Disabling automatic language detection (where available) has no effect.
The user is not located in a Welsh-speaking household or environment and has no personal use for the Welsh language.
When the app encounters speech it cannot recognise, it defaults to assuming the language must be Welsh rather than indicating uncertainty or providing an English fallback.
It always defaults to Welsh, and never to any other language.
Known Pattern:
This is not an isolated case. There are numerous reports—on public forums including Reddit—of users experiencing the exact same issue. The problem appears to disproportionately affect users in the UK, particularly those near Wales or using UK English system defaults. This suggests the model or the app may be overly influenced by regional metadata, or may have embedded assumptions about Welsh as a fallback language in certain contexts. I live in South Thailand, and Thai language is never a problem. In fact, it’s more intelligible.
Implications:
Several users have raised the possibility that this reflects not a simple bug but a structural decision—possibly influenced by government language support policies, or by models overcorrecting for inclusion of minority languages like Welsh. Whether or not that is the case, the result is the same: the user loses control, and the app fails to perform its basic function—accurate English speech recognition—on its default settings.
This raises not just usability concerns but political and cultural ones, particularly for users who are involuntarily subjected to a language they do not speak, in a context where precision and control are critical.
Requested Resolution:
A manual override in the ChatGPT app’s speech-to-text function to enforce English-only recognition, with no fallback, guessing, or substitution.
A clear option to disable multilingual detection entirely, so that users can rely on a predictable, consistent interface.
Public acknowledgment of the issue and its scope—particularly as it affects UK users.