Codex desktop on Windows: Czech voice input intermittently transcribed into unrelated languages/scripts

What variant of Codex are you using?

Codex desktop app on Windows.

Installed Windows package verified on the affected machine: OpenAI.Codex 26.903.9818.0.

What issue are you seeing?

Voice input intermittently produces text in an unrelated language or script during an otherwise Czech conversation. The user reports speaking clearly in Czech throughout, with no intentional language switching, but observes outputs resembling Slovak and Arabic on other occasions.

One concrete transcript observed in the affected conversation on September 12, 2026 is:

ਆਹ ਖਾਪੂ.

This example uses Gurmukhi characters. It is the actual text delivered to the assistant, not an inferred transcription or a translation added for this report. Surrounding user messages in the conversation are Czech. The assistant could not interpret the malformed message and asked the user to repeat it, interrupting the workflow.

Steps / observed sequence

1. Use voice input in the Codex desktop app on Windows during a conversation conducted in Czech.
2. Continue speaking Czech, including short follow-up messages.
3. Intermittently, the submitted transcription switches to an unrelated language/script rather than preserving Czech.
4. The assistant receives the incorrect text and cannot reliably recover the user's intended message.

This is a report of an observed incident and the user's recurring experience, not a deterministic minimal reproducer. The exact original Czech utterance for the example above and its audio recording are not available to the reporting assistant. The other reported language switches have not been independently verified from audio. The speech-recognition model and internal language-detection result are unknown.

Expected behavior

Spoken Czech should be transcribed faithfully into Czech without unsolicited translation or a switch to an unrelated script. An established Czech conversation should not repeatedly require the user to repair incorrectly detected languages. Please investigate language detection/context handling and consider a way to explicitly select or retain Czech for voice input.

Impact

Repeated clarification and re-dictation interrupts ordinary work and can change the meaning of instructions before they reach the assistant.

Related reports / scope

Related language/script-control report: #20003 (Traditional vs. Simplified Chinese; different language and platform). Searches for Czech voice input and wrong-language transcription did not identify an exact matching Czech report.

Only the minimal problematic transcript and environment version are included here. No audio, full conversation, workspace content, or diagnostic logs are attached.

Start by reproducing voice input in the Windows Codex desktop app during a Czech conversation, using the reported package version 26.903.9818.0, and inspect how the submitted transcript and language detection behave. Done means Czech speech remains Czech without unrelated scripts or unsolicited translation, or that the investigation identifies the responsible language-handling component.