Treat Narration Like Code: A Diff-Based TTS Workflow

One sentence changed.

A fictional habit-tracking app extended its trial from seven days to fourteen. Its three-minute onboarding narration was already approved, but this line was now wrong:

Your free trial lasts seven days.

Regenerating everything could change the pace or delivery of lines that did not need editing. Replacing only “seven days” might create an obvious join.

A cleaner fix is to divide the script into named sections and rebuild the affected sentence.

Give Each Line an ID

Store the script in a CSV file:

id,text,status
intro_01,"Welcome to the habit tracker.",approved
setup_01,"Choose one habit you want to build.",approved
trial_01,"Your free trial lasts seven days.",needs_update
finish_01,"You can change your reminder at any time.",approved

The revision becomes easy to see:

- trial_01,"Your free trial lasts seven days.",needs_update
+ trial_01,"Your free trial lasts fourteen days.",ready_for_generation

Use the same IDs for the audio files and add a version number:

intro_01_v1.wav
setup_01_v1.wav
trial_01_v2.wav
finish_01_v1.wav

A reviewer can now find the changed sentence without listening to the whole recording.

Replace the Sentence, Not Two Words

Generating only “fourteen days” may produce a different rhythm or volume. Regenerate the complete sentence instead:

Your free trial lasts fourteen days.

If the join still sounds rough, generate a slightly wider section with a neighbouring line for context. Then trim and place the approved sentence in a conventional audio editor.

This gives the new line some context without rebuilding the full narration.

Save the Settings and Version

Text is only part of the input. Keep an internal project record for the replacement. The example below is not an Index TTS API request:

{
  "segment_id": "trial_01",
  "script_version": "v3",
  "script_status": "approved_for_generation",
  "language": "English",
  "delivery": "clear and neutral",
  "speed": "normal",
  "reference": "approved_voice_01",
  "output": "trial_01_v3.wav",
  "audio_review": "pending"
}

This is more useful than a note saying “use the same voice as last time.”

Keep the earlier file until the new one is approved:

trial_01_v2.wav
trial_01_v3.wav

A reference can guide the voice, but separate generations may still differ in timing, tone or emphasis.

Use Text Differences to Guide the Update

The current Index TTS workflow accepts a script and an authorized reference recording. Depending on the selected model, the available controls may cover emotion, language, pronunciation, speed or duration. Its site also recommends working on longer projects in sections so individual lines remain easier to revise.

A simple update process looks like this:

1. Compare the old and new scripts.
2. Mark the changed line IDs.
3. Generate the affected sentences.
4. Listen to each replacement in context.
5. Approve and assemble the new version.

This is a production method, not a claim that separate outputs will sound identical. Check the current controls, costs and terms before starting.

Listen to the Join

A replacement may sound fine alone but wrong between two approved lines.

Check:

• Volume
• Speaking pace
• Pauses before and after the line
• Pronunciation of names, numbers and abbreviations
• Whether the new wording preserves the approved meaning

For this example, “fourteen” should not sound like “forty.” Test the assembled section on normal headphones or phone speakers, not only studio equipment.

Verify Facts Before Generating Speech

Natural delivery does not make a sentence correct.

The fourteen-day trial must come from an approved product record. Apply the same rule to prices, dates, features, names, safety information and regulated claims.

Approve the written script first. The voice tool should read the decision, not make it.

Check Permission for the Reference Voice

Before uploading a recording, confirm that the relevant agreement allows:

• External AI processing
• Voice cloning or use as a voice reference
• Derivative audio
• The intended publication or commercial reuse

Permission to publish the original recording does not automatically cover all of these uses.

Do not use meeting audio, podcast excerpts or a public figure’s voice simply because the recording is available. If listeners could mistake synthetic speech for a real statement by an identifiable person, clearly explain how the audio was created and where it is being used.

Keep Small Changes Small

A one-line update should not require rebuilding three minutes of narration.

Stable line IDs, saved settings and versioned files show what changed and what entered the final export.

Find the diff, rebuild the affected sentence and test it in context.