A pleasant greeting is a weak test for a narration system that will later have to read product names, dates and emotionally specific lines. Before comparing TTS outputs, define what the test sentence is meant to reveal—and what its correct reading should be.
This workflow adapts FablessBridge's published guide to building Japanese comparison scripts. It is a reusable review procedure, not a completed benchmark. The example sentences below have not been generated or scored for this article.
1. Build four small test cases
The source proposes four categories. Keep them relevant to the intended production rather than filling the script with unusually difficult words.
| Case | Illustrative Japanese input | What to inspect |
|---|---|---|
| Everyday sentence | 窓を閉めてから、こちらの席へお座りください。 | Phrase boundaries and natural pauses |
| Proper noun | 架空の商品「灯ノート」をご紹介します。 | Whether the fictional product name is read as 「ひのーと」 |
| Numbers | 受付は10月21日、午前9時からです。 | Date, time and missing numbers |
| Emotion | 間に合ってよかった。ずっと待っていたんです。 | Relief, and whether the acting becomes too strong |
These are explanatory fixtures from the source, not quoted production audio. Substitute the words and situations your application actually needs. Put exceptional edge cases in separate tasks so they do not silently replace your real use case.
2. Freeze the intended reading before listening
A name can have more than one plausible reading. Write down the intended one before judging the output, rather than changing the answer to fit what a model produced. For numbers, establish whether the context calls for individual digits, a quantity or a date.
Keep the display text separate from the text passed to speech generation. A pronunciation aid should not accidentally become the subtitle. Link the two fields with the same line ID.
For example, a minimal record for the fictional product-name case could be represented as:
{
"line_id": "proper_noun_01",
"display_text": "架空の商品「灯ノート」をご紹介します。",
"speech_text": "架空の商品「ひのーと」をご紹介します。",
"intended_name_reading": "ひのーと"
}
This JSON is an illustrative representation of the guide's fields, not an executed API request or a file from a measured run. The guide links to a separate 14-line production record where display and speech text were kept apart.
3. Record the conditions, not just the model name
The same sentence spoken with different voices, languages, speeds or performance instructions is not the same comparison condition. Keep those settings and the number of generations with the outputs. If providers do not offer the same voice, state that difference instead of implying it was controlled.
A review record should make it possible to answer:
- Which model and selected voice produced the file?
- What language, speed and performance instructions were used?
- Which input line and intended reading apply?
- How many generation attempts were made?
- Is this the original audio or a processed version?
4. Keep evaluation dimensions separate
Review textual accuracy, phrasing, voice quality and acting separately. A voice can be appealing while misreading the product name. A single overall score hides whether the next action is to correct the reading, adjust a pause or choose another voice.
The source does not prescribe a numerical scale or report a winning model. Use the categories to make the required correction identifiable, rather than presenting an unsupported ranking.
5. Compare originals first, then the finished mix
Retain the audio immediately after generation. If you adjust volume or edit pauses, record the processing and distinguish that version from the original during review.
Finally, check intelligibility with the actual subtitles and background music. An isolated voice test and the completed video answer different questions. These four sentence types are a starting point for your own application, not a comprehensive test of every speech capability.
A practical handoff for a narrated video
When commissioning a Japanese product video, provide official spellings and intended readings separately, especially for names and numbers. Agree on the required expression and retain the script and audio version being approved. This gives the editor a concrete basis for narration review.
Original Japanese guide · Related 14-line production record · FablessBridge production examples
For a video that combines narration, captions and music, discuss your brief or email fabless-bridge@shian.space.
Disclosure: This article was prepared with AI assistance from a published FablessBridge guide. No new TTS generation, scoring or model benchmark was performed for this post.
Top comments (0)