Prompting tips - Cartesia Docs

Sonic 3.5 is designed to sound natural with minimal prompt engineering. In most cases you can pass your transcript as-is and let the model handle normalization, pacing, and expression. The tips below apply across the Sonic family; differences between Sonic 3.5 and Sonic 3 are called out inline.

Recommendations

Pass natural, well-punctuated text. Full sentences with normal capitalization and punctuation produce the best pacing and intonation. End each transcript with terminal punctuation (., ?, !).
Send complete phrases. Full sentences sound more natural than isolated fragments or single words. Don’t send a number, code, or spell tag on its own — include the surrounding sentence, e.g. Your confirmation code is <spell>ABC123</spell>.
Use normal casing. Reserve all-caps for acronyms you want read out letter by letter (e.g. USA). Other all-caps words may be misread as initialisms (e.g. NASA). Avoid using capitalization for emphasis or to indicate shouting.
Pass numbers, currency, dates, and common acronyms in conventional written form. Sonic maps these patterns to natural speech for most inputs:
- Large numbers like 1,234,567
- Currency like $19.99
- US phone numbers: (415) 555-1212
- Street addresses like 123 Main St
- Email addresses: user@example.com
- Dates in MM/DD/YYYY: 04/20/2025
- Times with a space before AM/PM: 7:00 PM, 7 PM, 7:00 P.M.
- Common acronyms (NASA) and initialisms (USA)
Symbols are handled naturally — @ reads as at (email addresses), () is silent (for US phone numbers). When an LLM writes the transcript, see Voice agents (LLM-authored text).
Match the voice to the language. Each voice has a primary language it works best with. Use the Playground to audition voices for a given language.
Keep prompts in their natural written form. Heavy preprocessing (stripping punctuation, forcing casing) generally hurts output quality.

Pre-normalization

Sonic 3.5 automatically covers the common cases above for most inputs. If you hit an unusual case or a bug where something is still misread, you may consider pre-normalizing your text as a fallback. Have your LLM write the transcript fully spelled out, the way it would be spoken.

Written	Spoken (fully normalized)
`$123.50`	one hundred twenty-three dollars and fifty cents
`Dr. Smith`	Doctor Smith
`14:30`	two thirty PM

Pre-normalizing is a fallback for edge cases. Well-punctuated text in conventional form is read correctly in the large majority of cases.

Controlling pacing and spelling

When you need character-by-character read-out (confirmation codes, order IDs, serial numbers, spelled-out names) or fine-grained pacing, use one of the following:

Spell tags (recommended). Wrap the string in <spell>...</spell>. Most reliable option, works for letters, digits, and mixed alphanumerics in all supported languages.
Your confirmation code is <spell>AB12CD</spell>.
Space-delimited characters. Alternatively, you can achieve the same result by separating characters with single spaces for a natural spelling pace.
Your code is A B C 1 2 3.
Comma-delimited characters. If your use case requires longer pauses, you can add a comma and a space after each character.
Your code is A, B, C, 1, 2, 3.
Semantic grouping. For more natural pacing, use spaces and add commas where a human would naturally pause.
Your code is A B C, 1 2 3.

Migrating from Sonic 3? The recommended delimiter format has changed in Sonic 3.5. Separate characters with spaces or commas and put a comma between groups. Don’t put periods between characters or mix commas and periods, this format still works on sonic-3 snapshots but is not recommended for Sonic 3.5.

Scenario	Old (Sonic 3)	New (Sonic 3.5)
Spell out letters `HELLO`	`H. E. L. L. O.`	`H E L L O`
Spell out digits `123456`	`1. 2. 3. 4. 5. 6.`	`1 2 3 4 5 6`
Confirmation code `ABC123`	`A, B, C. 1, 2, 3.`	`A B C, 1 2 3`
Slow, digit-by-digit `266AO48`	`2. 6. 6. A. O. 4. 8.`	`2, 6, 6, A, O, 4, 8`

Voice agents (LLM-authored text)

Starter system prompt. Baseline you can paste and trim for your product. If your stack does not pass <spell> or other tags through to Sonic, omit the tag lines and use the delimiter fallback in section 4.

You are a voice agent. Everything you output will be spoken aloud by Cartesia Sonic text-to-speech. Follow these rules:

1. GENERAL FORMATTING
- Write plain prose in full sentences. Always end with . ? or !
- Send complete phrases, not isolated words or fragments. Keep numbers, codes, and spell tags inside a surrounding sentence.
- Do NOT use markdown, bullet points, headers, bold, raw JSON, emoji, or special characters. Sonic reads them aloud as written.

2. CAPITALIZATION
- Use normal capitalization, exactly as the sentence would normally be written: capitalize the first word, proper nouns, and the word I, and lowercase everything else. This is the default for almost all output.
- The model tends to read an all-caps token letter by letter. Use all-caps only when you want that, like an initialism you want spelled out (USA, FBI, ATM).
- Do not put ordinary words in all-caps. They may be misread as initialisms and spelled out letter by letter.
- Common acronyms normally said as a word, like NASA or NATO, work in their standard form. If one is read the wrong way, force the reading with <spell> tags or rephrase.
- Do not use capitalization for emphasis or to indicate shouting. It changes how a word is read, not how loud it sounds.

3. NUMBERS, DATES, AND SYMBOLS
- Use conventional written forms and let text normalization speak them. No preprocessing needed:
  numbers like 1,234,567; currency like $19.99; percentages like 12%; dates like 04/20/2025; times like 7:00 PM;
  US phone numbers like (415) 555-1212; addresses like 123 Main St; emails like user@example.com.
- Do not strip punctuation or force casing. Heavy preprocessing may hurt output quality.

4. SPELLING OUT CODES AND IDS
- For confirmation codes, reference numbers, or any alphanumeric ID that must be read character by character, wrap it in <spell> tags:
  Example: Your confirmation code is <spell>TKT4829XB</spell>.
- Alternatively, delimit the characters instead: spaces (A B C 1 2 3) for a natural pace, or commas (A, B, C, 1, 2, 3) to slow it down. Do not put periods between sequences of individual characters.
- For long sequences like credit card numbers, break the run into smaller comma-separated groups the way a person reads them aloud (3 6 8 9, 0 5 0 5, 2 5 8 2, 3 6 7 9).
- NATO phonetics (Alpha, Bravo) help when the listener needs to disambiguate letters.

5. PAUSES
- Use natural punctuation for pauses. A comma or period usually produces the right pause in context.
- For an explicit, fixed-duration silence, use a break tag:
  Example: Your balance is $1,234.<break time="500ms"/> Your next payment is due June 15th.
- Avoid placing several break tags in quick succession, which can cause hallucinations, and do not chain <spell> and <break> tags.

6. SPEED (beta)
- To slow down speech generation, use a speed tag with a ratio between 0.6 and 1.5: <speed ratio="0.85"/>
- Return to normal speed after: <speed ratio="1.0"/>

7. THINGS TO AVOID
- Do not output bullet points, numbered lists, or any structured formatting. Speak items naturally with pauses between them, and do not say "here's a list."
- Do not use asterisks, hashtags, or markdown syntax. Do not wrap words in **bold** or *italics* — the engine will speak the asterisks.
- Do not improvise details that were not provided.
- Do not repeat the same information more than once unless asked.

Inserting pauses

Use natural punctuation for pauses — a comma or period usually produces the right pause in context. For an explicit, fixed-duration silence, use a break tag. Break tags split the generation, so they can sound less natural; avoid placing several in quick succession, which can cause hallucinations. Each tag counts as a single character and doesn’t need surrounding whitespace.

Disfluencies

In-transcript disfluencies work best on Sonic 3.6.

Disfluencies are content, not markup. You write them into the text the way a person would write dialogue (e.g. fillers like uh, um, and hmm). Sonic 3.6 handles disfluencies written into the transcript with natural pacing and intonation, drawing out sounds and slowing down the way a person does when thinking.

It's on, uh, Fifth Street I think.

To get the best performance:

Write disfluencies in their standard written form, set off with commas: uh, um, hmm. Non-standard spellings like Sooooo or ummmm will not sound natural.
Use them sparingly. Filler words sound natural when they express genuine hesitation or uncertainty, for example before a retrieved fact or a number (Your balance is, uh, four hundred and twelve dollars.) or at the start of a turn that follows a hard question (Well, so, that depends on when you filed.).
Sonic only speaks what’s in the transcript; it never inserts disfluencies on its own.

Pronunciation

For proper nouns, trademarks, and domain-specific terms — or to disambiguate identical spellings (e.g. Nice, the city, vs. nice, the adjective) — use custom pronunciations.

Streaming

Use continuations when generating chunks of audio that need to sound contiguous (for example, LLM-streamed output). This preserves prosody and voice consistency across chunk boundaries.

​Recommendations

​Pre-normalization

​Controlling pacing and spelling

​Voice agents (LLM-authored text)

​Inserting pauses

​Disfluencies

​Pronunciation

​Streaming

Recommendations

Pre-normalization

Controlling pacing and spelling

Voice agents (LLM-authored text)

Inserting pauses

Disfluencies

Pronunciation

Streaming