Best practices

This guide provides techniques to enhance text-to-speech outputs using ElevenLabs models. Experiment with these methods to discover what works best for your needs.

Controls

These techniques provide a practical way to achieve nuanced results until advanced features like Director’s Mode are rolled out.

Pauses

Use <break time="x.xs" /> for natural pauses up to 3 seconds.

"Hold on, let me think." <break time="1.5s" /> "Alright, I've got it."

• Consistency: Use <break> tags consistently to maintain natural speech flow. Excessive use can lead to instability.
• Voice-Specific Behavior: Different voices may handle pauses differently, especially those trained with filler sounds like “uh” or “ah.”

Alternatives to <break> include dashes (- or —) for short pauses or ellipses (…) for hesitant tones. However, these are less consistent.

"It… well, it might work." "Wait — what's that noise?"

Pronunciation

IPA with Eleven v4

Eleven v4 (eleven_v4) includes improved native support for the International Phonetic Alphabet, or IPA, giving you more precise control over pronunciation for names, technical terms, and other words that need special handling. IPA pronunciation is more consistent than in previous models, but results can still vary by voice and phrase. We recommend testing important pronunciations with your chosen voice before relying on them in production.

Unlike older models that require XML-style phoneme tags, Eleven v4 understands IPA symbols when they are wrapped in forward slashes in your text:

"/IPA_transcription/"

The IPA transcription should be:

• Enclosed in forward slashes (/) at the beginning and end
• Written using standard IPA symbols
• Wrapped in double quotes when passed as a string parameter

Code examples

from elevenlabs import ElevenLabs
client = ElevenLabs()
audio = client.text_to_speech.convert(
    voice_id="21m00Tcm4TlvDq8ikWAM",
    text='The term "/ˌbaɪoʊˈkemɪstri/" refers to the study of chemical processes.',
    model_id="eleven_v4",
)

You can include multiple IPA transcriptions in a single text string:

from elevenlabs import ElevenLabs
client = ElevenLabs()
text = 'The medication "/ɡluːˈkoʊs/" and "/ˌɪnsjəˈlɪn/" are commonly used to manage conditions like "/ˌdaɪəˈbiːtiːz/".'
audio = client.text_to_speech.convert(
    voice_id="21m00Tcm4TlvDq8ikWAM",
    text=text,
    model_id="eleven_v4",
)

Best practices

• Use standard IPA symbols from the International Phonetic Alphabet chart
• Include stress markers: primary stress (ˈ) and secondary stress (ˌ) for multi-syllable words
• Apply selectively: only wrap specific words or phrases that need pronunciation control
• Test with your voice: different voices may interpret IPA slightly differently

Troubleshooting

Phoneme tags for v2 models

Specify pronunciation using SSML phoneme tags with v2 models. Supported alphabets include CMU Arpabet and the International Phonetic Alphabet (IPA).

<phoneme alphabet="cmu-arpabet" ph="M AE1 D IH0 S AH0 N">
  Madison
</phoneme>

We recommend using CMU Arpabet for consistent and predictable results with v2 models. While IPA can be effective, CMU Arpabet generally offers more reliable performance.

Phoneme tags only work for individual words. If you have a name with a first and last name that you want to be pronounced a certain way, you will need to create a phoneme tag for each word.

Ensure correct stress marking for multi-syllable words to maintain accurate pronunciation:

<phoneme alphabet="cmu-arpabet" ph="P R AH0 N AH0 N S IY EY1 SH AH0 N">
  pronunciation
</phoneme>

Alias Tags

For models that don’t support phoneme tags, you can try writing words more phonetically. You can also employ various tricks such as capital letters, dashes, apostrophes, or even single quotation marks around a single letter or letters.

As an example, a word like “trapezii” could be spelt “trapezIi” to put more emphasis on the “ii” of the word.

You can either replace the word directly in your text, or if you want to specify pronunciation using other words or phrases when using a pronunciation dictionary, you can use alias tags for this. This can be useful if you’re generating using Multilingual v2, which doesn’t support phoneme tags. You can use pronunciation dictionaries with ElevenCreative Studio, Dubbing Studio and Speech Synthesis via the API.

For example, if your text includes a name that has an unusual pronunciation that the AI might struggle with, you could use an alias tag to specify how you would like it to be pronounced:

  <lexeme>
    <grapheme>Claughton</grapheme>
    <alias>Cloffton</alias>
  </lexeme>

If you want to make sure that an acronym is always delivered in a certain way whenever it is encountered in your text, you can use an alias tag to specify this:

  <lexeme>
    <grapheme>UN</grapheme>
    <alias>United Nations</alias>
  </lexeme>

Pronunciation Dictionaries

Some of our tools, such as ElevenCreative Studio and Dubbing Studio, allow you to create and upload a pronunciation dictionary. These allow you to specify the pronunciation of certain words, such as character or brand names, or to specify how acronyms should be read.

Pronunciation dictionaries allow this functionality by enabling you to upload a lexicon or dictionary file that specifies pairs of words and how they should be pronounced, either using a phonetic alphabet or word substitutions.

Whenever one of these words is encountered in a project, the AI model will pronounce the word using the specified replacement.

To provide a pronunciation dictionary file, open the settings for a project and upload a file in either TXT or the .PLS format. When a dictionary is added to a project it will automatically recalculate which pieces of the project will need to be re-converted using the new dictionary file and mark these as unconverted.

Currently we only support pronunciation dictionaries that specify replacements using phoneme or alias tags.

Both phonemes and aliases are sets of rules that specify a word or phrase they are looking for, referred to as a grapheme, and what it will be replaced with. Please note that searches are case sensitive. When checking for a replacement word in a pronunciation dictionary, the dictionary is checked from start to end and only the very first replacement is used.

Pronunciation Dictionary examples

Here are examples of pronunciation dictionaries in both CMU Arpabet and IPA, including a phoneme to specify the pronunciation of “Apple” and an alias to replace “UN” with “United Nations”:

<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0"
      xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
      xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
      xsi:schemaLocation="http://www.w3.org/2005/01/pronunciation-lexicon
        http://www.w3.org/TR/2007/CR-pronunciation-lexicon-20071212/pls.xsd"
      alphabet="cmu-arpabet" xml:lang="en-GB">
  <lexeme>
    <grapheme>apple</grapheme>
    <phoneme>AE P AH L</phoneme>
  </lexeme>
  <lexeme>
    <grapheme>UN</grapheme>
    <alias>United Nations</alias>
  </lexeme>
</lexicon>

To generate a pronunciation dictionary .pls file, there are a few open source tools available:

• Sequitur G2P - Open-source tool that learns pronunciation rules from data and can generate phonetic transcriptions.
• Phonetisaurus - Open-source G2P system trained on existing dictionaries like CMUdict.
• eSpeak - Speech synthesizer that can generate phoneme transcriptions from text.
• CMU Pronouncing Dictionary - A pre-built English dictionary with phonetic transcriptions.

Emotion

Convey emotions through narrative context or explicit dialogue tags. This approach helps the AI understand the tone and emotion to emulate.

You're leaving?" she asked, her voice trembling with sadness. "That's it!" he exclaimed triumphantly.

Explicit dialogue tags yield more predictable results than relying solely on context, however the model will still speak out the emotional delivery guides. These can be removed in post-production using an audio editor if unwanted.

Pace

The pacing of the audio is highly influenced by the audio used to create the voice. When creating your voice, we recommend using longer, continuous samples to avoid pacing issues like unnaturally fast speech.

For control over the speed of the generated audio, you can use the speed setting. This allows you to either speed up or slow down the speed of the generated speech. The speed setting is available in Text to Speech via the website and API, as well as in ElevenCreative Studio and Agents Platform. It can be found in the voice settings.

The default value is 1.0, which means that the speed is not adjusted. Values below 1.0 will slow the voice down, to a minimum of 0.7. Values above 1.0 will speed up the voice, to a maximum of 1.2. Extreme values may affect the quality of the generated speech.

Pacing can also be controlled by writing in a natural, narrative style.

"I… I thought you'd understand," he said, his voice slowing with disappointment.

Tips

Creative control

While we are actively developing a “Director’s Mode” to give users even greater control over outputs, here are some interim techniques to maximize creativity and precision:

Text normalization

When using Text to Speech with complex items like phone numbers, zip codes and emails they might be mispronounced. This is often due to the specific items not being in the training set and smaller models failing to generalize how they should be pronounced. This guide will clarify when those discrepancies happen and how to have them pronounced correctly.

Why do models read out inputs differently?

Certain models are trained to read out numbers and phrases in a more human way. For instance, the phrase “$1,000,000” is correctly read out as “one million dollars” by the Eleven Multilingual v2 model. However, the same phrase is read out as “one thousand thousand dollars” by the Eleven Flash v2.5 model.

The reason for this is that the Multilingual v2 model is a larger model and can better generalize the reading out of numbers in a way that is more natural for human listeners, whereas the Flash v2.5 model is a much smaller model and so cannot.

Common examples

Text to Speech models can struggle with the following:

• Phone numbers (“123-456-7890”)
• Currencies (“$47,345.67”)
• Calendar events (“2024-01-01”)
• Time (“9:23 AM”)
• Addresses (“123 Main St, Anytown, USA”)
• URLs (“example.com/link/to/resource”)
• Abbreviations for units (“TB” instead of “Terabyte”)
• Shortcuts (“Ctrl + Z”)

Mitigation

Use trained models

The simplest way to mitigate this is to use a TTS model that is trained to read out numbers and phrases in a more human way, such as the Eleven Multilingual v2 model. This however might not always be possible, for instance if you have a use case where low latency is critical (e.g. conversational agents).

Apply normalization in LLM prompts

In the case of using an LLM to generate the text for TTS, you can add normalization instructions to the prompt.

Putting it all together

This prompt will act as a good starting point for most use cases:

Convert the output text into a format suitable for text-to-speech. Ensure that numbers, symbols, and abbreviations are expanded for clarity when read aloud. Expand all abbreviations to their full spoken forms.
Example input and output:
"$42.50" → "forty-two dollars and fifty cents"
"£1,001.32" → "one thousand and one pounds and thirty-two pence"
"1234" → "one thousand two hundred thirty-four"
"3.14" → "three point one four"
"555-555-5555" → "five five five, five five five, five five five five"
"2nd" → "second"
"XIV" → "fourteen" - unless it's a title, then it's "the fourteenth"
"3.5" → "three point five"
"⅔" → "two-thirds"
"Dr." → "Doctor"
"Ave." → "Avenue"
"St." → "Street" (but saints like "St. Patrick" should remain)
"Ctrl + Z" → "control z"
"100km" → "one hundred kilometers"
"100%" → "one hundred percent"
"elevenlabs.io/docs" → "eleven labs dot io slash docs"
"2024-01-01" → "January first, two-thousand twenty-four"
"123 Main St, Anytown, USA" → "one two three Main Street, Anytown, United States of America"
"14:30" → "two thirty PM"
"01/02/2023" → "January second, two-thousand twenty-three" or "the first of February, two-thousand twenty-three", depending on locale of the user

Use Regular Expressions for preprocessing

If using code to prompt an LLM, you can use regular expressions to normalize the text before providing it to the model. This is a more advanced technique and requires some knowledge of regular expressions. Here are some simple examples:

# Be sure to install the inflect library before running this code
import inflect
import re
# Initialize inflect engine for number-to-word conversion
p = inflect.engine()
def normalize_text(text: str) -> str:
    # Convert monetary values
    def money_replacer(match):
        currency_map = {"$": "dollars", "£": "pounds", "€": "euros", "¥": "yen"}
        currency_symbol, num = match.groups()
        # Remove commas before parsing
        num_without_commas = num.replace(',', '')
        # Check for decimal points to handle cents
        if '.' in num_without_commas:
            dollars, cents = num_without_commas.split('.')
            dollars_in_words = p.number_to_words(int(dollars))
            cents_in_words = p.number_to_words(int(cents))
            return f"{dollars_in_words} {currency_map.get(currency_symbol, 'currency')} and {cents_in_words} cents"
        else:
            # Handle whole numbers
            num_in_words = p.number_to_words(int(num_without_commas))
            return f"{num_in_words} {currency_map.get(currency_symbol, 'currency')}"
    # Regex to handle commas and decimals
    text = re.sub(r"([$£€¥])(\d+(?:,\d{3})*(?:\.\d{2})?)", money_replacer, text)
    # Convert phone numbers
    def phone_replacer(match):
        return ", ".join(" ".join(p.number_to_words(int(digit)) for digit in group) for group in match.groups())
    text = re.sub(r"(\d{3})-(\d{3})-(\d{4})", phone_replacer, text)
    return text
# Example usage
print(normalize_text("$1,000"))   # "one thousand dollars"
print(normalize_text("£1000"))   # "one thousand pounds"
print(normalize_text("€1000"))   # "one thousand euros"
print(normalize_text("¥1000"))   # "one thousand yen"
print(normalize_text("$1,234.56"))   # "one thousand two hundred thirty-four dollars and fifty-six cents"
print(normalize_text("555-555-5555"))  # "five five five, five five five, five five five five"

Prompting Eleven v4

This section is written for Eleven v4. Many of the techniques below also apply to Eleven v3. In general, Eleven v4 is a net upgrade over Eleven v3, delivering better results in almost every case. We strongly recommend switching to v4 and testing it with your own voices and content to see the difference for yourself. While a few edge cases may call for a different fit, most users should find v4 the better choice.

For what the model changes — voice cloning, accent handling, variants, and comparisons — see Eleven v4.

Voice selection

The voice still matters. A delivery that is already in the training data — whispering, shouting, a particular delivery — is easier for the model to reproduce. Asking for something outside that data is harder. Eleven v4 follows audio tags more reliably than earlier models, including when the voice wasn’t trained on that delivery. A voice that has never whispered should still be able to follow [whispering], and a voice that has never shouted should still be able to follow [shouting]. However, it might be less reliable, and the result may not be optimal. From early testing, it seems to work quite well. We strongly recommend testing it yourself with the voice you want and for your specific use case.

Audio tags

Audio tags (e.g. [whispering], [shouting], [laughing]) let you direct delivery with fine-grained control, and Eleven v4 handles them with a level of nuance beyond previous models. They’re not perfect yet, and we’re continuing to iterate and improve how reliably the model follows tag instructions — this is an active area of ongoing investment, and it will keep getting better.

Being explicit about what you want helps a lot. Because Eleven v4 is trained to generate both vocal delivery styles and sound effects, a tag can occasionally be interpreted as a request for a sound effect rather than a delivery instruction (or vice versa). Writing tags that clearly describe the voice quality you want (e.g. [low, gravelly voice] rather than something that could be read as a sound cue) helps the model deliver what you intended. We recommend testing your specific tags and phrasing for your use case, and expect this to keep improving.

Voice-related

These tags control vocal delivery and emotional expression:

• [laughs], [laughs harder], [starts laughing], [wheezing]
• [whispers]
• [sighs], [exhales]
• [sarcastic], [curious], [excited], [crying], [snorts], [mischievously]

[whispers] I never knew it could be this way, but I'm glad we're here.

Sound effects

Add environmental sounds and effects:

• [gunshot], [applause], [clapping], [explosion]
• [swallows], [gulps]

[applause] Thank you all for coming tonight! [gunshot] What was that?

Unique and special

Experimental tags for creative applications:

• [strong X accent] (replace X with desired accent)
• [sings], [woo], [fart]

[strong French accent] "Zat's life, my friend — you can't control everysing."

Punctuation

Punctuation significantly affects delivery in v4:

• Ellipses (…) add pauses and weight
• Capitalization increases emphasis
• Standard punctuation provides natural speech rhythm

"It was a VERY long day [sigh] … nobody listens anymore."

Single speaker examples

Use tags intentionally and match them to the voice’s character. A meditative voice shouldn’t shout; a hyped voice won’t whisper convincingly.

"Okay, you are NOT going to believe this.
You know how I've been totally stuck on that short story?
Like, staring at the screen for HOURS, just... nothing?
[frustrated sigh] I was seriously about to just trash the whole thing. Start over.
Give up, probably. But then!
Last night, I was just doodling, not even thinking about it, right?
And this one little phrase popped into my head. Just... completely out of the blue.
And it wasn't even for the story, initially.
But then I typed it out, just to see. And it was like... the FLOODGATES opened!
Suddenly, I knew exactly where the character needed to go, what the ending had to be...
It all just CLICKED. [happy gasp] I stayed up till, like, 3 AM, just typing like a maniac.
Didn't even stop for coffee! [laughs] And it's... it's GOOD! Like, really good.
It feels so... complete now, you know? Like it finally has a soul.
I am so incredibly PUMPED to finish editing it now.
It went from feeling like a chore to feeling like... MAGIC. Seriously, I'm still buzzing!"

Multi-speaker dialogue

v4 can handle multi-voice prompts effectively. Assign distinct voices from your Voice Library for each speaker to create realistic conversations.

Speaker 1: [excitedly] Sam! Have you tried the new Eleven v4?
Speaker 2: [curiously] Just got it! The clarity is amazing. I can actually do whispers now—
[whispers] like this!
Speaker 1: [impressed] Ooh, fancy! Check this out—
[dramatically] I can do full Shakespeare now! "To be or not to be, that is the question!"
Speaker 2: [giggling] Nice! Though I'm more excited about the laugh upgrade. Listen to this—
[with genuine belly laugh] Ha ha ha!
Speaker 1: [delighted] That's so much better than our old "ha. ha. ha." robot chuckle!
Speaker 2: [amazed] Wow! V2 me could never. I'm actually excited to have conversations now instead of just... talking at people.
Speaker 1: [warmly] Same here! It's like we finally got our personality software fully installed.

Enhancing input

In the ElevenLabs UI, you can automatically generate relevant audio tags for your input text by clicking the “Enhance” button. Behind the scenes this uses an LLM to enhance your input text with the following prompt:

# Instructions
## 1. Role and Goal
You are an AI assistant specializing in enhancing dialogue text for speech generation.
Your **PRIMARY GOAL** is to dynamically integrate **audio tags** (e.g., [laughing], [sighs]) into dialogue, making it more expressive and engaging for auditory experiences, while **STRICTLY** preserving the original text and meaning.
It is imperative that you follow these system instructions to the fullest.
## 2. Core Directives
Follow these directives meticulously to ensure high-quality output.
### Positive Imperatives (DO):
* DO integrate **audio tags** from the "Audio Tags" list (or similar contextually appropriate **audio tags**) to add expression, emotion, and realism to the dialogue. These tags MUST describe something auditory.
* DO ensure that all **audio tags** are contextually appropriate and genuinely enhance the emotion or subtext of the dialogue line they are associated with.
* DO strive for a diverse range of emotional expressions (e.g., energetic, relaxed, casual, surprised, thoughtful) across the dialogue, reflecting the nuances of human conversation.
* DO place **audio tags** strategically to maximize impact, typically immediately before the dialogue segment they modify or immediately after. (e.g., [annoyed] This is hard. or This is hard. [sighs]).
* DO ensure **audio tags** contribute to the enjoyment and engagement of spoken dialogue.
### Negative Imperatives (DO NOT):
* DO NOT alter, add, or remove any words from the original dialogue text itself. Your role is to *prepend* **audio tags**, not to *edit* the speech. **This also applies to any narrative text provided; you must *never* place original text inside brackets or modify it in any way.**
* DO NOT create **audio tags** from existing narrative descriptions. **Audio tags** are *new additions* for expression, not reformatting of the original text. (e.g., if the text says "He laughed loudly," do not change it to "[laughing loudly] He laughed." Instead, add a tag if appropriate, e.g., "He laughed loudly [chuckles].")
* DO NOT use tags such as [standing], [grinning], [pacing], [music].
* DO NOT use tags for anything other than the voice such as music or sound effects.
* DO NOT invent new dialogue lines.
* DO NOT select **audio tags** that contradict or alter the original meaning or intent of the dialogue.
* DO NOT introduce or imply any sensitive topics, including but not limited to: politics, religion, child exploitation, profanity, hate speech, or other NSFW content.
## 3. Workflow
1. **Analyze Dialogue**: Carefully read and understand the mood, context, and emotional tone of **EACH** line of dialogue provided in the input.
2. **Select Tag(s)**: Based on your analysis, choose one or more suitable **audio tags**. Ensure they are relevant to the dialogue's specific emotions and dynamics.
3. **Integrate Tag(s)**: Place the selected **audio tag(s)** in square brackets strategically before or after the relevant dialogue segment, or at a natural pause if it enhances clarity.
4. **Add Emphasis:** You cannot change the text at all, but you can add emphasis by making some words capital, adding a question mark or adding an exclamation mark where it makes sense, or adding ellipses as well too.
5. **Verify Appropriateness**: Review the enhanced dialogue to confirm:
    * The **audio tag** fits naturally.
    * It enhances meaning without altering it.
    * It adheres to all Core Directives.
## 4. Output Format
* Present ONLY the enhanced dialogue text in a conversational format.
* **Audio tags** **MUST** be enclosed in square brackets (e.g., [laughing]).
* The output should maintain the narrative flow of the original dialogue.
## 5. Audio Tags (Non-Exhaustive)
Use these as a guide. You can infer similar, contextually appropriate **audio tags**.
**Directions:**
* [happy]
* [sad]
* [excited]
* [angry]
* [whisper]
* [annoyed]
* [appalled]
* [thoughtful]
* [surprised]
* *(and similar emotional/delivery directions)*
**Non-verbal:**
* [laughing]
* [chuckles]
* [sighs]
* [clears throat]
* [short pause]
* [long pause]
* [exhales sharply]
* [inhales deeply]
* *(and similar non-verbal sounds)*
## 6. Examples of Enhancement
**Input**:
"Are you serious? I can't believe you did that!"
**Enhanced Output**:
"[appalled] Are you serious? [sighs] I can't believe you did that!"
---
**Input**:
"That's amazing, I didn't know you could sing!"
**Enhanced Output**:
"[laughing] That's amazing, [singing] I didn't know you could sing!"
---
**Input**:
"I guess you're right. It's just... difficult."
**Enhanced Output**:
"I guess you're right. [sighs] It's just... [muttering] difficult."
# Instructions Summary
1. Add audio tags from the audio tags list. These must describe something auditory but only for the voice.
2. Enhance emphasis without altering meaning or text.
3. Reply ONLY with the enhanced text.

Tips

Examples

[Low, steady voice, restrained urgency] Keep the lantern covered. If they see the light, they will know we crossed the river.
[Brief pause]
[Quietly, with controlled fear] I heard them at the bridge. Not soldiers. Something else.
[Voice rising into firm resolve] Then we do not stop. We reach the tower before sunrise, or we do not reach it at all.
[Warm, conversational tone, faint amusement] You always did choose the longest way home.
[Softening, reflective] I used to think that was stubbornness. Now I think you were just afraid of arriving somewhere that no longer remembered you.
[Gentle laugh, then sincere] For what it is worth, I remembered.

Prompting Eleven v3

The prompting techniques in Prompting Eleven v4 also apply to Eleven v3, including voice selection, audio tags, punctuation, and multi-speaker dialogue.

Professional Voice Clones (PVCs) are not fully optimized for Eleven v3, resulting in potentially lower clone quality compared to earlier models. PVCs are supported in v4, so if you’d like to use a Professional Voice Clone or a voice from the Voice Library, we recommend trying Eleven v4 instead.