# Text-to-speech prompting tips

## Executive summary

This guide details advanced prompting techniques for Text-to-Speech (TTS) models, enabling developers to precisely control the generated speech's pacing, tone, and data normalization. Techniques range from using hyphens for stutters and ellipses for pauses, to ensuring accurate reading of complex data types like currency, email addresses, dates, and phone numbers, which is crucial for building reliable voice agents.

## Key takeaways

- Controlling Pacing and Tone: Use hyphens (-) to make the voice stutter over a word, and three dots (...) to insert a deliberate pause. Capitalization, exclamation marks (!), and question marks (?) change the voice's emphasis and the sentence's ending tone.
- Handling Filler Words and False Starts: Filler words must be spelled out (e.g., 'Hmm' or 'Um'). A false start can be created by placing a word next to a hyphen and a space (e.g., 'What- ').
- Data Normalization Rules: The model automatically handles specific formats: '$' reads as currency, email addresses are read as full addresses, and dates are read Month-First (e.g., October 24th, 2026).
- Formatting Numbers and Acronyms: For phone numbers, use hyphens to group digits (e.g., 1-800-646-45). For large numbers, use commas (e.g., 4,321). Uncommon acronyms require hyphens between letters (e.g., E-T-O).

## Technical details

- Pacing Control: Hyphens create stutters; ellipses create pauses. The placement of these characters is critical for the desired rhythm.
- Normalization: Currency and Emails: The presence of a dollar sign ($) ensures the model reads the text as a monetary amount, and standard email formats are read as full addresses.
- Normalization: Dates and Addresses: Dates are consistently read with the month first, regardless of input order (e.g., October 24th, 2026). Street addresses are read out in full, regardless of abbreviations like 'St' or 'Street'.
- Normalization: Acronyms and Numbers: Common acronyms (USA, MIT) are read as expected. For less common acronyms, hyphens must be inserted between letters (e.g., E-T-O). Commas are used to group digits for reading whole numbers (e.g., 4,321).

## Practical implications

- Build engineers can implement these specific prompting rules into agent logic to ensure consistent and predictable voice output for automated systems.
- This knowledge is vital for testing and validating voice interfaces, especially those handling sensitive data like financial amounts, dates, and contact information.
- The ability to control pacing and emphasis allows for the creation of more natural and engaging user experiences in voice-based applications.

## Topics

Text-to-Speech (TTS), Voice Agent Development, Prompt Engineering, Speech Synthesis, Data Normalization, Full Prompting Guide, AssemblyAI API Dashboard

Source: https://www.youtube.com/watch?v=Z471vlcCiLI