TTS Best Practices

This document provides practical guidance for using SESTEK Text-to-Speech effectively. It explains how general text formatting, punctuation, spacing, special characters, and output configuration influence TTS behavior, and it offers best practices for improving pronunciation, pacing, prosody, and synthesis stability across common use cases.

By following these guidelines, you can produce more natural-sounding output and better understand expected system behavior when diagnosing undesired synthesis results.

1. Purpose and Scope

Objectives

• Enable higher-quality TTS outputs
• Explain how the TTS system interprets general written text
• Reduce ambiguity when diagnosing synthesis issues
• Provide clear guidance for text preparation and request formatting
• Improve output naturalness through better punctuation, spacing, and prosody control

Scope

• General Text-to-Speech (TTS) best practices
• Text preparation before synthesis
• Special character handling
• Request formatting and JSON escaping
• Output-oriented troubleshooting

2. TTS Processing Overview

SESTEK TTS follows this synthesis pipeline:

Input Text
    ↓
Text Normalization
    ↓
Prosody Assignment
    ↓
Phoneme Generation
    ↓
Waveform Synthesis

What This Means in Practice

Even when the words themselves are correct, the following may still affect the final result:

• missing punctuation
• excessive punctuation
• incorrect spacing
• unresolved placeholders
• raw special characters
• malformed JSON payloads
• ambiguous numeric or symbolic expressions
• unsuitable sentence length or structure

3. Prosody and Why It Matters

In TTS, prosody refers to how speech is delivered, including:

• Pause placement
• Pause length
• Rhythm
• Phrase grouping
• Emphasis
• Sentence melody / intonation

Punctuation and special characters influence prosody even when they are not spoken literally.

Examples:

• A comma usually creates a short pause
• A period usually creates a sentence-ending pause
• A question mark usually creates question-like intonation
• An ellipsis may create a hesitation or trailing pause
• Repeated symbols may create unnatural rhythm or exaggerated emphasis

4. Special Characters Reference Tables

Special characters may affect synthesis in different ways. Some characters are explicitly handled by the SESTEK TTS engine as pause delimiters with configured pause durations.

Other characters are not defined as engine-level pause delimiters, but may still affect pronunciation, rhythm, readability, or output naturalness depending on the text and context.

The pause durations in the first table are configured at engine level and are expressed in milliseconds. Actual perceived pause length may still vary depending on the voice, model behavior, and surrounding text.

4.1. SESTEK TTS Pause Delimiter Characters

The table below lists the characters currently supported by the SESTEK TTS engine as pause delimiters.

4.2. Other Common Characters and Their Possible Effects

The table below lists other common characters that are not defined as engine-level pause delimiters, but may still affect synthesis depending on context.

The “Relative Audio Representation” column is illustrative only. Actual output may vary depending on voice, language, model behavior, normalization rules, and surrounding text.

4.3. Prosody-Sensitive Character Combinations

Some character combinations may produce different prosodic effects than the individual characters alone. This is especially important when punctuation or delimiter characters are preserved from display-oriented text and passed directly to TTS.

Adjacent delimiter characters may create stronger, less natural, or unexpected prosodic boundaries depending on the configured pause durations and how the text is interpreted during processing.

This is particularly relevant when sentence punctuation or list separators are combined with delimiter characters such as quotation marks or vertical bars.

Examples:

Key Rules

• Validate spoken naturalness as well as text correctness.
• Remove unnecessary quotation marks from enumerations or category lists where possible.
• Avoid combining delimiter characters with commas unless the result has been validated in synthesis output.
• Prefer punctuation that supports natural prosody over display-style quoting.

5. Text Preparation and Input Cleaning

High-quality TTS output starts with clean and well-prepared input text.

Recommended

• Use complete and readable sentences
• Keep punctuation intentional and minimal
• Use single spacing between words
• Resolve placeholders before synthesis
• Remove raw HTML, XML, JSON fragments, and template markers
• Rewrite technical strings into spoken-friendly text when possible

Avoid

• repeated punctuation
• inconsistent spacing
• raw variable placeholders
• markup tags
• decorative symbols
• code-like strings in customer-facing prompts

Examples:

6. Spacing and Sentence Structure

6.1. Word Spacing

Spacing errors frequently lead to poor synthesis quality.
Tips:

• Use a single space between words
• Keep punctuation attached correctly to the preceding word
• Avoid overloading one sentence with too much information

Examples:

6.2. Sentence Length

Very long sentences may reduce naturalness and make phrasing unstable.

Tips:

• Split long paragraphs into shorter sentences

Examples:

7. Numbers, Symbols, and Spoken-Friendly Rewriting

Symbols and compressed written forms do not always produce the most natural spoken output.

Tips:

• Prefer the spoken form whenever clarity matters more than visual compactness.

Examples:

8. URLs, Emails, and Technical Strings

8.1. URLs and Technical Strings

URLs, file paths, usernames, and technical identifiers often sound unnatural when synthesized directly.

Tips:

• Only keep such content in TTS input if the system is expected to read it literally.

Examples:

8.2. Emails

Emails may be normalized to spoken forms automatically, but exact pronunciation of handles, punctuation, and domains may vary.

Tips:

• Keep the email address only if it must be spoken literally.
• Otherwise, rewrite it as a spoken instruction.
• Define abbreviations for domains.
• Separate confusing punctuation marks (dots, hyphens, underscore, etc.) at the end of the email address.

Examples:

Tips:

• Rewrite the sentence
• insert a pause using space character or SSML if supported
• or express the email address in spoken form.

Examples (English):

Examples (Turkish):

9. Pronunciation and Language Customization

9.1. Abbreviations

Abbreviation management allows custom pronunciations for abbreviations and acronyms. Default system behavior applies unless explicitly customized.

For advanced control, refer to: Customization for Synthesis Accuracy

Tips:

• Define abbreviations or custom pronunciations for foreign-origin words and brand names that do not have clear equivalents in the target language or script.
• Decide whether the abbreviation should be spoken:
  • as a full word
  • letter-by-letter
  • or as its expanded form
• Prefer the representation that matches the intended spoken output.

Examples:

9.2. Names, Brand Names, Foreign-Origin Words, and Mixed-Language Text

Proper nouns, brand names, product names, and foreign-origin words are a common source of pronunciation variation.

Mixed-language text may also reduce naturalness, especially when scripts or pronunciation rules change within the same sentence.

Tips:

• Define abbreviations for foreign-origin words, brand names that have no equivalent in the target language.
• Prefer the form that is closest to the intended spoken output.
• Expand abbreviations letter-by-letter or segment-by-segment when needed in context.

Examples (Turkish):

Examples (English):

9.3. Normalization

Normalization converts written text into a spoken-friendly representation before synthesis. This helps improve pronunciation of numbers, dates, and other complex text elements.

9.3.1. Numbers

Numeric values are automatically converted to spoken forms. Context determines whether numbers are read digit-by-digit or as whole values.

Long numeric sequences can cause numbers to sound too fast or dense.

Tips:

• Write numbers in the exact representation form of their value for natural synthesis.
• Rewrite the number in words when a full-value reading is required.
• Separate digits when digit-by-digit reading is intended.
• Insert pauses for long identifiers or codes.
• For additional pause guidance, see Section 10.

9.3.2. Dates

Numeric and alphanumeric date forms are automatically converted to spoken forms. Month phrases are always spoken even when input is in numeric format.

Tips:

• Prefer unambiguous date formats.
• Use month names when clarity matters.
• Keep separators and formatting consistent.

9.3.3. Times

Numeric and alphanumeric time forms are automatically converted to spoken forms. Spoken output depends on detected format and surrounding context.

Tips:

• Prefer clear and standard time formats.
• Rewrite unusual or compressed forms when clarity matters.

Examples

9.3.4. Currencies

Currency names and symbols are normalized to spoken forms automatically.

Tips:

• Prefer full currency names for correct synthesis.

Examples:

• Define abbreviations for currency codes if their spoken form must be controlled.

Examples (English):

Examples (Turkish):

9.3.5. Addresses

There is no address-specific normalization. Numeric parts in addresses follow standard number normalization rules.

Province, street, and building names, flat numbers, and dense address strings may be interpreted unexpectedly. Long numeric sequences may sound dense without pauses.

Tips:

• Add commas at natural phrase boundaries.
• Expand abbreviations where needed.
• Separate long numeric segments clearly.

Examples:

9.3.6. Symbols

The pronunciation of symbols may be interpreted differently depending on language and surrounding context.

Tips:

• Define abbreviations for symbols for customized pronunciation.

Examples (Turkish):

• Symbol: "&" → Default pronunciation: "ve".

10. Pauses and Flow Control

10.1. Punctuation

Punctuation influences pauses automatically, but manual pause control may still be useful in specific cases.

Tips:

• Use punctuation first for natural phrasing
• Use manual pause control only when necessary
• Place pauses at semantic boundaries
• Avoid excessive pause insertion in short text

Examples:

• Without structure:
  Your verification code is 1 2 5 0 7 8 and your order number is 4 5 2 1 please keep them for reference.
• Better structured:
  Your verification code is 1 2 5 0 7 8. Your order number is 4 5 2 1. Please keep them for reference.

10.2. SSML Tags

SSML tags allow customizing pronunciation, intonation, and emphasis. SESTEK supports the most commonly used SSML tags including <break>, <say-as>, <voice>, and audio insertion.
For full details, refer to: SSML Tag Support.

Tips:

• Use pauses sparingly - avoid excessive breaks in short sentences
• Place pauses at semantic boundaries
• For long numeric sequences, group digits (3–3–4 or 2–2–2) and add short breaks

Examples:

• Without SSML pauses:
  Your verification code is 1 2 5 0 7 8 and your order number is 4 5 2 1 please keep them for reference.

• With SSML pauses:

<speak>Your verification code is<break time='300ms'/> 1 2 5<break time='250ms'/>0 7 8<break time='250ms'/> and your order number is<break time='300ms'/>4 5 2 1<break time='300ms'/>please keep them for reference</speak>

11. Why Output May Sound Unexpected

Common Causes

• Missing punctuation
• Excessive punctuation
• Unresolved placeholders
• Raw special characters
• Dense numeric content
• Long or poorly segmented sentences
• Raw URLs or technical strings
• Channel/output mismatch
• Malformed payloads caused by incorrect JSON escaping

Examples:

12. Voice Selection, Rate, Volume, and Supported Controls

12.1. Voices

Different voices may vary in clarity, pacing, and handling of numbers, acronyms, or foreign-origin words.
For the full list of supported voices, refer to: Supported Languages and Voices

12.2. Adjustable Parameters

Depending on the deployment and selected voice, the following parameters may be available:

• Rate - controls the speaking tempo of the voice.
• Volume - controls the base loudness level of the voice.

Recommended Approach
Start with default values, then adjust gradually and validate with representative prompts.

12.3. Emotion, Tone, and Phoneme Tags

Explicit emotion or tone controls and phoneme tags are not supported.
Emotional delivery cannot be directly controlled via markup. Indirect influence is possible through:

• Sentence structure
• Punctuation
• Wording
• Strategic pause placement

For implementation-specific markup support, refer to: SSML Tag Support

13. Audio Output Format and Playback Compatibility Guide

• Match synthesis output settings to the target playback channel
• Avoid unnecessary transcoding where possible
• Use telephony-appropriate settings for IVR/call-center playback
• Validate synthesized audio in the actual downstream environment
• Be aware that playback issues may originate from channel limitations rather than the synthesis itself

14. Troubleshooting Guide

A word, abbreviation, or brand name is pronounced incorrectly

Possible causes: Undefined abbreviation, acronym ambiguity, foreign-origin word, mixed-language text, or unexpected default normalization.
Actions: Define a custom pronunciation if supported, rewrite the term in a spoken-friendly form, expand the abbreviation, or keep the sentence in one language/script where possible.

Numbers, dates, or times sound unexpected

Possible causes: Ambiguous format, dense numeric content, insufficient pauses, or a whole-number vs. digit-by-digit mismatch.
Actions: Rewrite the value in the intended spoken form, separate digits when needed, use unambiguous date/time formats, and add pauses where clarity is critical.

Output does not sound as intended in mixed-language content

Possible causes: Frequent language-switching, foreign-origin words, or inconsistent script usage within the same sentence.
Actions: Keep the sentence in one language where possible, use localized spellings/transliterations in language-specific deployments, or define recurring foreign terms through abbreviation.

Output sounds too fast or dense

Possible causes: Long sentences, dense numeric blocks, insufficient punctuation.
Actions: Split the text into shorter sentences, add punctuation, group content more clearly.

Output sounds robotic or overly segmented

Possible causes: Too much punctuation, excessive commas, too many forced pauses.
Actions: Simplify punctuation and remove unnecessary separators.

Symbols are spoken literally

Possible causes: Raw symbols were sent directly in the input.
Actions: Rewrite symbols in spoken form where needed.

Placeholders or markup are spoken

Possible causes: Template variables or tags were not resolved before synthesis.
Actions: Clean and normalize the text before sending it to TTS.

Request fails before audio is generated

Possible causes: Invalid JSON, especially unescaped " or \.
Actions: Validate and escape the payload correctly.

15. Summary

Recommended:

• Use clean, spoken-friendly text
• Apply punctuation intentionally
• Keep spacing consistent
• Rewrite symbols and compact forms when clarity matters
• Remove placeholders, markup, and technical artifacts
• Escape JSON payloads correctly
• Use unambiguous number, date, time, and currency formats
• Define abbreviations or custom pronunciations for recurring acronyms, names, and foreign-origin terms
• Keep one language/script per sentence where possible
• Match output settings to the target playback environment
• Use sentence structure to support natural prosody
• Validate voice, rate, and volume settings with representative prompts

Avoid:

• Repeated decorative punctuation
• Meaningless Prosodic Grouping.
• Raw URLs and code-like text in customer-facing prompts
• Leaving unresolved placeholders in synthesis input
• Assuming every symbol will be read naturally
• Diagnosing voice quality before validating input formatting
• Treating screen-friendly formatting as automatically speech-friendly
• Relying on ambiguous abbreviations without customization
• Mixing languages/scripts within the same sentence unnecessarily
• Expecting emotion control tags
• Passing raw numeric-heavy text without preparation

16. Final Note

Well-formed TTS output starts with well-prepared input. In many cases, improving punctuation, spacing, symbol usage, and request formatting is enough to significantly improve synthesis clarity, prosody, and stability without changing the voice itself.