Make Japanese TTS Revisions Traceable: Separate Captions, Readings and Audio Files

A pronunciation correction should have a clear scope: which line changes, which audio file is replaced, and which subtitle must stay intact.

In a FablessBridge production run on October 16, 2026, we generated 14 Japanese narration lines as separate files with Gemini 3.8 Flash TTS. The saved files total approximately 44.72 seconds. That is evidence of the saved outputs and their durations, not a benchmark of voice quality or an end-to-end video runtime.

1. Keep two text fields linked by one line ID

Use one version for the subtitle and another for speech. The subtitle keeps the intended spelling, kanji and official product names. The speech version can express difficult readings in hiragana or katakana.

For example, a fictional announcement such as “新サービスの受付は10月21日からです” can keep its written date on screen while the speech text specifies its pronunciation. This is an explanatory example, not a quotation from the recorded narration.

The important invariant is the line ID: changing the reading must not silently replace the caption. Review both fields together when the meaning changes; review only the reading field when correcting pronunciation.

2. Make the line the regeneration unit

Generate and save each line separately, then assemble the files in the editor. A correction can target that line instead of regenerating the entire narration.

This does not guarantee a seamless result. Short files can differ at their boundaries in breath, volume or delivery. Review the corrected file with the preceding and following lines, not only in isolation.

3. Hold the voice instructions constant

The documented run used the model ID gemini-3.8-flash-tts and a shared direction: calm advertising narration, natural Japanese, restrained intonation and a slightly brisk pace. It received WAV output and converted that to MP3 for editing.

Keep the original WAV alongside the editing copy. When something sounds wrong, compare the original, converted file and assembled timeline. This helps distinguish a generated-output problem from a conversion or editing problem. Changing the model, direction and text at once makes that distinction harder.

4. Validate the handoff before accepting a revision

A reusable review sequence is:

1. Match the requested correction to its line ID.
2. Check that the speech text contains the intended reading and the caption still has the approved spelling.
3. Listen for omitted words, names, numbers and sentence endings.
4. Join the line to its neighbors and check pauses and volume transitions.
5. Review the sequence with the actual visuals, captions and BGM.
6. Keep the approved file and its original output so later edits have a clear reference.

The source records 14 saved files and their combined duration. It does not claim that these checks produced a quantified improvement in intelligibility. The run also had a $0.50 cap and no automatic retries; that cap is a workflow limit, not a quoted production price or a verified invoice total.

What this changes in a production brief

If you are commissioning a narrated product video, provide official spellings and intended readings separately. Ask how a one-word correction will be identified, replaced and checked in the final mix. This gives the review a concrete object instead of a request to “make the voice better.”

Original production record · FablessBridge production examples

For a service explainer that needs coordinated narration, captions and motion, discuss your brief or email [e-mail].

Disclosure: This article was drafted with AI assistance and adapted from FablessBridge's published production record. It reports that record rather than a new independent model test.