SSML Complete Guide: Control AI Speech Like a Pro (2026)

SSML (Speech Synthesis Markup Language) is the standard way to control how text-to-speech engines pronounce and deliver your content. Instead of flat, robotic output, SSML gives you fine-grained control over:

• Pronunciation — fix how specific words sound
• Pacing — speed up or slow down parts of your audio
• Volume — emphasize words or whisper them
• Pitch — raise or lower intonation
• Pauses — add silence for dramatic effect
• Breaths — insert natural breathing sounds

This guide covers every major SSML tag with working examples. Most of these work with Google Cloud TTS, Azure Speech, Amazon Polly, and ElevenLabs.

Quick Start

SSML wraps your text in <speak> tags:

<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis">
  Hello, this is AI speech with SSML control.
</speak>

To use it with any major TTS API:

# Google Cloud TTS
from google.cloud import texttospeech
client = texttospeech.TextToSpeechClient()
ssml = '<speak>Hello <break time="500ms"/> world.</speak>'
synthesis_input = texttospeech.SynthesisInput(ssml=ssml)

# ElevenLabs
from elevenlabs import generate
audio = generate(text='<speak>Hello world.</speak>'),  # ElevenLabs auto-detects SSML

Complete Tag Reference

<break> — Pauses and Silence

The most commonly used SSML tag. Controls silence between words.

<speak>
  No pause.
  <break time="200ms"/> Short pause.
  <break time="1s"/> One second pause.
  <break strength="weak"/> Weak paragraph break.
  <break strength="strong"/> Strong paragraph break.
  <break strength="x-strong"/> Extra strong break.
</speak>

Strength values (approximate durations):

strength	Typical pause
none	0ms
x-weak	250ms
weak	500ms
medium	750ms
strong	1000ms
x-strong	1500ms

Pro tip: Use strength when the exact timing doesn’t matter — the engine will choose a natural duration. Use time="500ms" when you need precise, repeatable timing (e.g., for video synchronization).

---

<prosody> — Pitch, Rate, and Volume

Controls the musical qualities of speech.

<speak>
  <!-- Speech rate: slower and faster -->
  <prosody rate="slow">This is slow speech.</prosody>
  <prosody rate="x-slow">Very slow.</prosody>
  <prosody rate="fast">This is fast speech.</prosody>
  <prosody rate="x-fast">Very fast.</prosody>
  <prosody rate="80%">80% of normal speed.</prosody>
  <prosody rate="150%">50% faster.</prosody>

  <!-- Pitch: higher and lower -->
  <prosody pitch="low">This has low pitch.</prosody>
  <prosody pitch="high">This has high pitch.</prosody>
  <prosody pitch="-20%">20% lower pitch.</prosody>
  <prosody pitch="+30%">30% higher pitch.</prosody>

  <!-- Volume: quieter and louder -->
  <prosody volume="silent">
    <break time="1s"/>
  </prosody>
  <prosody volume="x-soft">Very quiet.</prosody>
  <prosody volume="soft">Quiet speech.</prosody>
  <prosody volume="medium">Normal volume.</prosody>
  <prosody volume="loud">Loud speech.</prosody>
  <prosody volume="x-loud">Very loud.</prosody>
  <prosody volume="-6dB">6 decibels quieter.</prosody>

  <!-- Combined: all three -->
  <prosody rate="slow" pitch="+10%" volume="loud">
    Combined attributes for dramatic narration.
  </prosody>
</speak>

Real-world example — narration with natural pacing:

<speak>
  <prosody rate="slow" pitch="low">
    It was a dark and stormy night.
  </prosody>
  <break time="500ms"/>
  <prosody rate="medium" volume="loud">
    Suddenly, the door burst open!
  </prosody>
  <break time="300ms"/>
  <prosody rate="x-fast" pitch="high">
    She ran inside, breathless.
  </prosody>
</speak>

---

<emphasis> — Word-Level Stress

Marks words that should stand out.

<speak>
  I <emphasis level="moderate">really</emphasis> mean it.
  I <emphasis level="strong">absolutely</emphasis> mean it.
  This is the <emphasis level="reduced">least important</emphasis> part.
</speak>

Level	Effect
strong	Maximum emphasis, higher pitch and volume
moderate	Default, noticeable but natural
reduced	De-emphasized, lower and quieter
none	No emphasis

---

<say-as> — Interpret Text Correctly

Controls how numbers, dates, and abbreviations are pronounced.

<speak>
  <!-- Numbers -->
  <say-as interpret-as="cardinal">42</say-as>  <!-- "forty-two" -->
  <say-as interpret-as="ordinal">42</say-as>    <!-- "forty-second" -->
  <say-as interpret-as="digits">42</say-as>     <!-- "four two" -->

  <!-- Date formats -->
  <say-as interpret-as="date" format="ymd">2026-07-24</say-as>  <!-- "July 24th, 2026" -->
  <say-as interpret-as="date" format="mdy">07/24/2026</say-as>
  <say-as interpret-as="date" format="dmy">24/07/2026</say-as>

  <!-- Characters -->
  <say-as interpret-as="characters">HTML</say-as>  <!-- "H T M L" -->
  <say-as interpret-as="spell-out">AI</say-as>     <!-- "A I" -->

  <!-- Telephone numbers -->
  <say-as interpret-as="telephone">555-0123</say-as>

  <!-- Fractions -->
  <say-as interpret-as="fraction">1/2</say-as>  <!-- "one half" -->
  <say-as interpret-as="fraction">3/4</say-as>  <!-- "three quarters" -->
</speak>

---

<phoneme> — Fix Pronunciation

The most powerful tag for accuracy. Uses IPA (International Phonetic Alphabet) to specify exact pronunciation.

<speak>
  <!-- Fixing common mispronunciations -->
  I enjoy eating <phoneme alphabet="ipa" ph="ˈtoʊ.mɑː.təʊ">tomato</phoneme>.
  The <phoneme alphabet="ipa" ph="əˈskeɪ.dʒəs">esophagus</phoneme> connects the throat to the stomach.
  The company <phoneme alphabet="ipa" ph="ˈliː.noʊ">Leno</phoneme> was founded in 2020.
  <phoneme alphabet="ipa" ph="ˈniː.kɒn">Nikon</phoneme> cameras are excellent.

  <!-- Using Google's x-sampa phonetic alphabet (alternative) -->
  <phoneme alphabet="x-sampa" ph="'[email protected]">tomato</phoneme>
</speak>

Pro tip: Google’s Phoneme Visualizer is invaluable for finding correct IPA transcriptions. Also try ipa-reader to test pronunciations.

---

<audio> — Insert Sound Effects

Available in Amazon Polly and some other providers. Inserts audio files or sound effects into speech.

<speak>
  Welcome to our podcast!
  <audio src="https://example.com/intro-music.mp3">
    <break time="2s"/>
  </audio>
  Today we're discussing AI voice technology.
</speak>

The text inside <audio> is used as fallback if the audio file can’t be loaded.

---

<p> and <s> — Paragraph and Sentence Boundaries

Explicitly marks structural units for better prosody.

<speak>
  <p>
    <s>This is the first sentence of the first paragraph.</s>
    <s>This is the second sentence.</s>
  </p>
  <p>
    <s>This is the first sentence of a new paragraph.</s>
  </p>
</speak>

---

<sub> — Substitution

Replace displayed text with different spoken text.

<speak>
  The <sub alias="World Health Organization">WHO</sub> issued new guidelines.
  We support <sub alias="Artificial Intelligence">AI</sub> research.
  Open <sub alias="Monday through Friday">Mon-Fri</sub>.
</speak>

---

<lang> — Language Switching

Switch between languages within a single SSML document.

<speak>
  The French word <lang xml:lang="fr">bonjour</lang> means hello.
  In Spanish, <lang xml:lang="es">gracias</lang> means thank you.
  <lang xml:lang="de">Guten Morgen</lang> is German for good morning.
</speak>

---

<par> and <media> — Parallel Audio (Azure Only)

Azure Speech supports parallel audio streams.

<speak version="1.0" xmlns:mstts="http://www.w3.org/2001/mstts">
  <par>
    <media begin="0s">
      <audio src="https://example.com/background.wav"/>
    </media>
    <media begin="0s">
      Welcome to this presentation!
    </media>
  </par>
</speak>

---

Provider-Specific Tags

Azure Speech — Expressiveness and Style

<speak version="1.0" xmlns:mstts="http://www.w3.org/2001/mstts"
       xmlns:emo="http://www.w3.org/2009/10/emotionml">
  <!-- Speaking style -->
  <mstts:express-as style="cheerful">
    Great news! Our project is going live.
  </mstts:express-as>

  <mstts:express-as style="sad">
    We regret to inform you...
  </mstts:express-as>

  <mstts:express-as style="whisper" styledegree="1.5">
    This is a secret message.
  </mstts:express-as>

  <!-- Available Azure styles: cheerful, sad, angry, fearful,
       excited, friendly, hopeful, shouting, whispering,
       terrified, unfriendly, whispering, cold, embarrassed -->
</speak>

ElevenLabs — SSML Support

ElevenLabs supports a subset of SSML including <break>, <prosody>, <phoneme>, <say-as>, and <emphasis>. Their newer models handle SSML tags more naturally than older ones.

---

Real-World Examples

Audiobook Narration

<speak>
  <prosody rate="medium" pitch="-5%">
    Chapter Three: The Discovery
  </prosody>
  <break time="1s"/>
  <prosody rate="slow">
    The morning sun <break time="200ms"/> cast long shadows across the room.
  </prosody>
  <prosody rate="medium" volume="loud" pitch="+10%">
    "There you are!" <break time="150ms"/> she exclaimed.
  </prosody>
  <prosody rate="medium">
    He turned slowly, <break time="300ms"/>
    his face illuminated by the pale light.
  </prosody>
</speak>

E-Learning Narration

<speak>
  <prosody rate="medium">
    Welcome to Module 4: <break time="200ms"/>
    Machine Learning Fundamentals.
  </prosody>
  <break time="500ms"/>
  <prosody rate="80%">
    First, let's understand the key concept.
    <break time="300ms"/>
    Machine learning is a <emphasis level="strong">subset</emphasis>
    of artificial intelligence that enables systems
    to <say-as interpret-as="characters">AI</say-as> to learn from data.
  </prosody>
  <break time="400ms"/>
  <prosody rate="90%">
    Important: <break time="200ms"/>
    <emphasis level="strong">Always validate your training data</emphasis>
    before starting the model training process.
  </prosody>
</speak>

YouTube Voice-Over

<speak>
  <prosody rate="fast" pitch="+10%">
    Hey everyone, welcome back to the channel!
    <break time="400ms"/>
  </prosody>
  <prosody rate="medium">
    Today we're reviewing the <sub alias="Text to Speech">TTS</sub>
    landscape in 2026.
    <break time="200ms"/>
    There are <say-as interpret-as="cardinal">11</say-as>
    major providers to choose from.
  </prosody>
</speak>

Testing Your SSML

Most providers offer SSML preview tools:

• Google Cloud: Text-to-Speech SSML tester
• Azure Speech: Audio Content Creation Studio
• Amazon Polly: AWS Console TTS tester
• ElevenLabs: Paste SSML directly into the ElevenLabs Speech Synthesis

Common Pitfalls

Mistake	Why	Fix
Missing namespace	SSML won’t parse	Always include xmlns="http://www.w3.org/2001/10/synthesis"
Self-closing breaks	<break/> may not work in all providers	Always use <break time="500ms"/> with explicit attribute
Over-nesting	Some providers limit tag depth	Keep SSML flat — max 3-4 levels of nesting
Wrong phonetic alphabet	IPA works everywhere, x-sampa is Google-only	Stick to IPA (alphabet="ipa") for cross-provider compatibility
Forgetting encoding	Special characters break parsing	Use XML entities: &amp; for &, &lt; for <, &gt; for >

Compatibility Matrix

Tag	Google Cloud	Azure	Amazon Polly	ElevenLabs
<break>	✅ Full	✅ Full	✅ Full	✅ Full
<prosody>	✅ Full	✅ Full	✅ Full	✅ Partial
<emphasis>	✅	✅	✅	✅
<say-as>	✅	✅	✅	✅
<phoneme>	✅ IPA/x-sampa	✅ IPA/SAPI	✅ IPA/x-sampa	✅ IPA
<sub>	✅	✅	✅	❌
<p>/<s>	✅	✅	✅	❌
<lang>	✅	✅	✅	❌
<audio>	❌	✅	✅	❌
<par>/<media>	❌	✅	❌	❌
mstts:express-as	❌	✅	❌	❌
<voice>	✅	❌	❌	❌

Bottom Line

SSML is the difference between robotic speech and professional-quality AI voice output. Even basic tags — <break>, <prosody>, and <say-as> — dramatically improve naturalness. For production content (audiobooks, e-learning, YouTube voice-overs), SSML is not optional — it’s the standard.