Codex Desktop voice dictation intermittently outputs Simplified Chinese for Traditional Chinese users

What variant of Codex are you using?

Codex Desktop App on macOS.

Related local CLI version, for reference:

codex-cli 0.125.0

What issue are you seeing?

Codex Desktop built-in voice dictation frequently transcribes spoken Traditional Chinese / Taiwan Mandarin into Simplified Chinese.

This makes the voice input workflow unusable for Traditional Chinese users, especially in projects or workspaces where all communication, code comments, documentation, and user instructions must remain in Traditional Chinese.

The important point is that this happens before the message is submitted to the agent. System, developer, or AGENTS.md instructions can force the agent's reply to use Traditional Chinese, but they cannot reliably fix the raw voice transcript already inserted into the composer.

Steps to reproduce

1. Open Codex Desktop App on macOS.
2. Use the built-in voice input / dictation feature in the composer.
3. Speak Mandarin intended to be written in Traditional Chinese, for example Taiwan usage.
4. Observe the text inserted into the composer.

Expected behavior

Codex should provide a way to force voice dictation output to Traditional Chinese, ideally Taiwan Traditional Chinese:

• zh-Hant
• zh-Hant-TW
• Traditional Chinese (Taiwan)

When the user has configured Traditional Chinese in system/developer/project instructions, Codex should not insert Simplified Chinese into the composer for Chinese voice dictation.

Actual behavior

The transcript intermittently uses Simplified Chinese. In practice this happens often enough that the built-in voice input is not usable for Traditional Chinese workflows.

Examples of undesired output include Simplified Chinese terms or characters being inserted into the composer even though the workspace instructions require Traditional Chinese.

Why this matters

For Traditional Chinese users, this is not a minor formatting preference. Simplified and Traditional Chinese are different written standards, and automatic post-hoc conversion is not always a safe 1:1 mapping. Some Simplified characters map to multiple possible Traditional characters depending on context.

This also breaks workflows that rely on strict project instructions such as:

• all replies must be in Traditional Chinese
• all generated comments and documentation must use Taiwan terminology
• if Simplified Chinese appears in the input, it should be treated as a voice transcription error

Those instructions only help after the agent receives the text. They do not fix the composer transcript itself.

Requested fix

Please add one or more of the following:

1. A voice dictation language/script setting for Chinese, including zh-Hant-TW.
2. A transcription prompt or style hint specifically for Codex Desktop dictation.
3. A post-transcription normalization layer before inserting text into the composer.
4. An option to automatically normalize Chinese voice transcripts to Traditional Chinese (Taiwan), similar to OpenCC s2twp behavior.

Related context

This appears related to the general limitation where Chinese transcription may use a single zh language bucket and produce either Simplified or Traditional output depending on model behavior. However, Codex Desktop currently does not expose a practical control to users who require Traditional Chinese.

Related issues:

• #12883 requested configurable voice transcription instructions.
• #13714 requested post-transcription cleanup for voice-dictated drafts.

This issue is narrower: Traditional Chinese users need reliable script control for Codex Desktop voice dictation.

Additional information

Local environment:

macOS
codex-cli 0.125.0

I can provide repeated examples or screenshots if needed.