Language versions of a project

flowchart TB
  A[Reviewed Polish source checkpoint] --> B[German session]
  A --> C[Japanese session]
  B --> D[Translation review and German settings]
  C --> E[Translation review and Japanese settings]
  D --> F[Speech plan and German export]
  E --> G[Speech plan and Japanese export]

Use a translation project when one recording needs several language versions. Choose a reviewed source checkpoint first. Each target language has its own session, settings, speech plan and outputs; it is not just a different export filename.

A Polish source with German and Japanese outputs

1. Transcribe the Polish clip and review names, numbers and speaker changes. If you already have reliable subtitles, import them instead.
2. Correct the text and select the checkpoint that the translations should share.
3. Add German and Japanese language versions. Choose suitable translation and speech models for each.
4. Review each translation before preparing its speech plan. Use language-matched voice references where cloning is available.
5. Check caption line lengths, reading speed and timing in each language. Generate a short voiceover sample, then assemble and export the selected branch.

Changing the shared source checkpoint does not silently refresh existing language sessions. Decide which branches should be rebuilt and review their existing changes first.

Subtitle limits belong to the target

By default, a new branch uses its target language's subtitle defaults. The optional carry_source_subtitle_settings choice copies the source's effective subtitle settings at branch creation.

For example, carrying a source limit of 42 characters and 15 characters per second keeps those values in the new Japanese branch. Without that option, Japanese defaults apply. Later changes to global defaults do not rewrite the copied values. Check whether carrying the source layout is appropriate for the translation.

Recognition, alignment and speech are separate

A model can recognize a language without providing supported word timing for it. Qwen3-ASR recognition includes Polish, while its native forced aligner's language list is narrower. Pandrator can use a supported timing fallback for some languages; this is not available for every recognized language.

Likewise, a multilingual speech model may require a language-specific voice or reference. Choose that language explicitly and listen to a sample. Automatic language detection and a successful job are not substitutes for checking pronunciation, missing words and timing.

Exports and automation

Confirm the selected audio run and written caption revision before each export. Display captions may use the current written text or the text captured with a generation run. Pronunciation changes made only for speech should not leak into ordinary captions.

Through MCP, inspect the translation project's source checkpoint and each branch's returned session/work handles. Follow those existing handles until completion. See caption workflows and the subtitle pipeline reference.