SSML for Beginners: Make Text-to-Speech Sound Human (Without Over-Engineering)

You've picked a great TTS voice, but something still sounds off. The pronunciation of your product name is wrong. The pause between sentences feels rushed. A question sounds like a statement.

SSML (Speech Synthesis Markup Language) is how you fix these problems. It's a standard markup language that lets you control exactly how text-to-speech engines speak your content—pronunciation, pacing, emphasis, and more.

The good news: you don't need to learn the whole spec. A handful of tags solve 90% of real-world TTS issues.

The tags you'll actually use

1. <break> — Add pauses

The most useful tag by far. Insert pauses where you want the voice to breathe:

<speak>
  Welcome to the demo. <break time="500ms"/> Let's get started.
</speak>

Use time to specify duration (milliseconds or seconds) or strength for semantic pauses (weak, medium, strong).

When to use it:

• Between sentences that run together
• Before important information
• After questions (to let them land)

2. <emphasis> — Stress words

Tell the engine which words deserve emphasis:

<speak>
  This is <emphasis level="strong">really</emphasis> important.
</speak>

Levels: reduced, moderate, strong

When to use it:

• Key terms or names
• Correcting unexpected stress patterns
• Making comparisons clear ("not X, but Y")

3. <say-as> — Pronounce things correctly

Handles formats that TTS engines otherwise mangle:

<speak>
  Call us at <say-as interpret-as="telephone">8005551234</say-as>.
  The meeting is on <say-as interpret-as="date">2025-03-15</say-as>.
</speak>

Useful interpret-as values: telephone, date, time, currency, characters (spell out), ordinal, cardinal

When to use it:

• Phone numbers
• Dates and times
• Abbreviations that should be spelled out
• Numbers that need specific formatting

4. <phoneme> — Override pronunciation

When the engine gets a word completely wrong, spell out how to say it:

<speak>
  Welcome to <phoneme alphabet="ipa" ph="ˈtɑːkioʊ">Talkio</phoneme>.
</speak>

When to use it:

• Brand names
• Technical terms
• Names from other languages
• Anything the engine consistently mispronounces

5. <prosody> — Control pitch, rate, and volume

Prosody—the stress, rhythm, and intonation of speech—is what makes voices sound natural. The `` tag lets you adjust it:

<speak>
  <prosody rate="slow" pitch="+10%">This part is spoken slowly and slightly higher.</prosody>
</speak>

Attributes: rate, pitch, volume (with values like slow, fast, percentages, or semitones)

When to use it:

• Slowing down for complex information
• Speeding up for repetitive content
• Adjusting energy level

For a deeper understanding of why prosody matters so much, see our piece on stress, rhythm, and AI voice quality.

A practical workflow

Don't mark up everything. Start with plain text and add SSML only where something sounds wrong:

1. Generate audio with plain text
2. Listen for problems (mispronunciations, rushed transitions, flat emphasis)
3. Add tags only where needed
4. Test again and iterate

Over-engineering with SSML makes content hard to maintain. Use the lightest touch that solves the problem.

Common mistakes to avoid

• Forgetting the <speak> wrapper: Most engines require your SSML to be wrapped in <speak> tags
• Using tags your engine doesn't support: Not all TTS providers support all SSML tags—check your provider's documentation
• Going overboard: Heavy markup makes content brittle and hard to update
• Ignoring punctuation: Often, just adding a comma or period fixes timing issues without needing SSML

For choosing which TTS provider to use in the first place, see our evaluation guide for product teams.