Numbers, dates and prices in TTS: write the spoken form first

The method

Rewrite anything that is ambiguous aloud into the words you want heard, then send that text. Sume TTS takes a plain transcript, not SSML, so the text you write is the text that is read. "Q4 revenue rose 12% to $3.4M on 10/04" should become "fourth-quarter revenue rose twelve percent to three point four million dollars on October fourth".

This costs nothing extra: the spoken form is longer, so it adds a few characters, at $0.0475 per 1,000. Fifty extra characters is $0.002375.

What to rewrite

These patterns cause the most surprise. Test each on your own voice, because voices and languages differ.

A test sentence

Before you generate a long script, build a short sentence that includes one of each pattern you use, send it, and listen. Keep the sentence and rerun it when you change voice or model. If the voice misreads a brand or a person's name, add a pronunciation dictionary and pass its id as pronunciation_dict_id, so the fix applies to every job. See the guide on pronunciation for the details.

curl -X POST https://api.sume.com/v1/tts-1.0/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: numbers-test-001" \
  -d '{"transcript":"Fourth-quarter revenue rose twelve percent to three point four million dollars on October fourth.",
       "voice":{"id":"voi_YOUR_VOICE_ID"},"language":"en"}'

Keep one style guide

Write your choices down: how you read percentages, dates, currencies and acronyms. A one-page style guide means every writer on the team prepares scripts the same way, and every episode sounds consistent. Pair it with a frozen voice, speed and volume, so only the words change from one episode to the next.

• Spell foreign names phonetically if a dictionary entry is not enough.
• Use commas for short pauses and full stops for longer ones.
• Do not rely on a symbol being read; write the word.
• Regenerate only the sentence that changed.

A worked rewrite

Take the line: "Order #4821 ships 10/05 for $19.99 (was $24.99), save 20%." Written for the ear it becomes: "Order number four eight two one ships October fifth for nineteen ninety-nine, down from twenty-four ninety-nine. Save twenty percent." The second version is 45 characters longer, which is 0.045 x $0.0475 = $0.0021 more on a single read, and nobody has to guess how the voice will handle the hash sign or the parentheses.

Do this in a small script if you produce many lines: keep a map of patterns to spoken forms and apply it before you send. Review the output by reading it aloud once, then submit. The check takes a minute and saves the cost of regenerating a long file, because a misread price in a finished ad is far more expensive than any characters you add.

If a number must be read digit by digit, such as a phone number or order id, write the digits with spaces. If it must be read as a whole, spell it out as words. Choose per case, write the choice in your style guide, and apply it every time.