tts

name	tts
description	Convert text to speech using ElevenLabs with emotional audio tags, pronunciation control, and blog-to-audio workflows

ElevenLabs Text-to-Speech Skill

Generate natural, expressive speech from text using ElevenLabs. This skill covers voice selection, audio tags, pronunciation control, text normalization, and multi-speaker dialogue.

When to Use This Skill

• Converting blog posts or articles to audio
• Creating voiceovers for videos
• Generating podcast intros/outros
• Building conversational AI responses
• Producing audiobook chapters
• Creating voice demos or samples

Quick Reference

Model Selection

Model	Best For	Latency	Quality
Eleven v3	Expressive content, emotional range, audio tags	Medium	Highest
Multilingual v2	Multiple languages, number/currency handling	Medium	High
Turbo v2.5	Low-latency applications, conversational AI	Fast	Good
Flash v2	Speed-critical applications	Fastest	Acceptable

Default choice: Eleven v3 for most content. Multilingual v2 if you need reliable number pronunciation without preprocessing.

Stability Settings (v3 Only)

• Creative (0.0-0.3): Most expressive, responds well to audio tags, may hallucinate
• Natural (0.4-0.6): Balanced, closest to original voice
• Robust (0.7-1.0): Most stable, less responsive to direction

Default: Natural (0.5) for most use cases.

---

Core Techniques

1. Pauses

For v2 models: Use SSML break tags.

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

Maximum pause: 3 seconds. Don't overuse. Too many break tags cause instability.

For v3 models: SSML breaks don't work. Use these alternatives:

• Ellipses (...) for hesitant pauses
• Dashes (-- or —) for short pauses
• [short pause] or [long pause] audio tags
• Line breaks in text

"I thought about it for a while... and I realized you were right."

2. Pronunciation

Phoneme tags (v2 models only):

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

Use CMU Arpabet over IPA. More consistent results.

For v3 or when phonemes don't work:

Write words phonetically. Use capitals for emphasis, apostrophes for syllable breaks.

• "trapezii" → "trapezIi" (emphasize the "ii")
• "Nguyen" → "Win" or "Nwin"
• "GIF" → "jiff" or "giff" depending on your preference

Alias tags (for pronunciation dictionaries):

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

3. Emotion and Delivery

Narrative context works better than hoping the AI "gets it":

"You're leaving?" she asked, her voice trembling with sadness.

The model reads the emotional guidance. Remove it in post-production if you don't want it in the final audio.

For v3, use audio tags instead:

[sad] "You're leaving?"

---

Eleven v3 Audio Tags

v3 introduces audio tags for emotional control. These go in square brackets before or after the text they modify.

Voice/Emotion Tags

[laughs], [laughs harder], [starts laughing], [wheezing]
[whispers]
[sighs], [exhales]
[sarcastic], [curious], [excited], [crying], [snorts], [mischievously]
[happy], [sad], [angry], [annoyed], [appalled], [thoughtful], [surprised]
[professional], [sympathetic], [reassuring], [dramatic]

Sound Effect Tags

[gunshot], [applause], [clapping], [explosion]
[swallows], [gulps]
[short pause], [long pause]

Special Tags

[strong French accent]
[strong Russian accent]
[sings]
[whispers]

Replace "French" or "Russian" with any accent. Results vary by voice.

Tag Placement Examples

Before the line:

[excited] "You won't believe what just happened!"

After the line:

"I can't believe this is happening..." [sighs]

Mid-sentence:

"I thought about it [long pause] and I realized you were right."

Combined:

[whispers] "Don't tell anyone, but..." [giggles] "I already knew."

What Works, What Doesn't

Tags must match the voice's character. A calm, meditative voice won't shout convincingly. A hyped voice won't whisper well.

Works:

• Subtle emotional shifts
• Laughs, sighs, pauses
• Accent shifts (if voice has range)

Doesn't work:

• Asking a whisper voice to shout
• Requesting music or complex sound effects
• Tags that contradict the voice's training data

---

Text Normalization

Smaller models struggle with numbers, dates, currencies, and special formats. Fix this before sending to the API.

Problem Inputs

Input	Problem
$1,000,000	May say "one thousand thousand dollars"
555-123-4567	May read as one big number
2024-01-15	May read digits individually
9:30 AM	May stumble on format
example.com/path	Unpredictable URL handling

Solutions

Option 1: Use Multilingual v2. It handles these better natively.

Option 2: Preprocess text before sending to API.

Normalization Reference

Input	Normalized Output
$42.50	forty-two dollars and fifty cents
£1,001.32	one thousand and one pounds and thirty-two pence
555-555-5555	five five five, five five five, five five five five
2nd	second
3.14	three point one four
100%	one hundred percent
Dr.	Doctor
St.	Street (unless "St. Patrick")
Ctrl + Z	control z
14:30	two thirty PM

Python Normalization Function

import inflect
import re

p = inflect.engine()

def normalize_for_tts(text: str) -> str:
    # Money
    def money_replacer(match):
        currency_map = {"$": "dollars", "£": "pounds", "€": "euros", "¥": "yen"}
        symbol, num = match.groups()
        num_clean = num.replace(',', '')
        
        if '.' in num_clean:
            dollars, cents = num_clean.split('.')
            return f"{p.number_to_words(int(dollars))} {currency_map.get(symbol, 'currency')} and {p.number_to_words(int(cents))} cents"
        return f"{p.number_to_words(int(num_clean))} {currency_map.get(symbol, 'currency')}"
    
    text = re.sub(r"([$£€¥])(\d+(?:,\d{3})*(?:\.\d{2})?)", money_replacer, text)
    
    # Phone numbers
    def phone_replacer(match):
        return ", ".join(" ".join(p.number_to_words((d))  d  g)  g  .groups())
    
    text = re.sub(, phone_replacer, text)
    
     text

---

Multi-Speaker Dialogue

v3 handles multiple voices in a single generation. Assign distinct voices from your library.

Format

Speaker 1: [excited] "Did you see the news?"

Speaker 2: [curious] "No, what happened?"

Speaker 1: "They announced the new feature we've been waiting for!"

Speaker 2: [surprised] "Already? I thought that was months away."

Tips for Dialogue

1. Consistent speaker labels. Use "Speaker 1" or character names consistently.
2. Vary emotional tags. Don't make every line dramatic.
3. Natural interruptions. Use em-dashes for cut-offs: "I was just going to—"
4. Overlapping timing. v3 can handle speakers cutting each other off.

---

Content-Specific Templates

Blog Post to Audio

[conversational] Today I want to talk about something that's been bugging me for a while.

WordPress caching. [sighs] It's one of those topics that sounds simple but gets complicated fast.

Here's what actually matters... [short pause] your server response time and how efficiently your pages get built. Everything else? Noise.

Tutorial Narration

[professional] Step one. Open your WordPress dashboard and navigate to Plugins.

[reassuring] Don't worry if your screen looks slightly different. The core layout is the same across most themes.

Click "Add New" in the top left corner. [short pause] You should see a search bar appear.

Product Review

[thoughtful] FlyingPress has become my go-to caching plugin. Let me explain why.

[matter-of-fact] First, it's fast. Not "marketing fast" where they cherry-pick benchmarks. Actually fast. I tested it against WP Rocket on twelve sites. FlyingPress won every time.

[honest] Now, the downsides. [sighs] Documentation is sparse. If you're the type who needs hand-holding, this might not be for you.

Newsletter Read

[warm] Hey, welcome to The Friday Drop. I'm Gaurav, and this is issue one fifty-two.

[excited] This week we've got some big news in the WordPress world. [short pause] Plus, I finally tested that new hosting stack I've been curious about.

Let's get into it.

---

Voice Selection Guide

For Your Content Types

Content	Voice Characteristics	Stability
Blog posts	Conversational, warm, slight enthusiasm	Natural (0.5)
Tutorials	Clear, professional, patient	Natural-Robust (0.6)
Reviews	Honest, thoughtful, occasional skepticism	Natural (0.5)
Newsletter	Friendly, energetic, personal	Creative-Natural (0.4)
Technical docs	Neutral, clear, consistent	Robust (0.7)

Voice Library Collections

ElevenLabs maintains curated collections. For v3 specifically, check their "Voices for V3" collection. These are optimized for the new model's capabilities.

Creating Custom Voices (IVC)

For instant voice clones:

1. Emotionally diverse samples: Include both neutral and expressive recordings
2. Consistent audio quality: Same microphone, same room, same distance
3. Minimum 1 minute: More is better, up to 10 minutes
4. Clear speech: No background noise, no music, no overlapping speakers

---

Common Issues and Fixes

Issue: Unnatural pacing

Cause: Voice was trained on short clips or inconsistent samples.

Fix: Use speed setting (0.7-1.2 range). Or rewrite text with more natural rhythm. Add pauses with ellipses.

Issue: Wrong emphasis

Cause: Model doesn't understand which words matter.

Fix: CAPITALIZE important words. Or restructure sentence to put key info at the end.

Issue: Pronunciation errors

Cause: Unusual names, technical terms, or foreign words.

Fix: Write phonetically, use alias tags, or create a pronunciation dictionary.

Issue: Audio tags ignored

Cause: Using Robust stability, or tag contradicts voice character.

Fix: Lower stability to Natural or Creative. Or choose a different voice that matches the desired emotion.

Issue: Hallucinations (random sounds, repeated words)

Cause: Stability too low, or text is ambiguous.

Fix: Increase stability. Simplify text structure. Remove conflicting emotional cues.

---

API Integration Notes

Basic Request Structure

import requests

url = "https://api.elevenlabs.io/v1/text-to-speech/{voice_id}"

headers = {
    "xi-api-key": "your-api-key",
    "Content-Type": "application/json"
}

data = {
    "text": "[excited] This is a test of the text to speech system.",
    "model_id": "eleven_v3",
    "voice_settings": {
        "stability": 0.5,
        "similarity_boost": 0.75,
        "speed": 1.0
    }
}

response = requests.post(url, json=data, headers=headers)

with open("output.mp3", "wb") as f:
    f.write(response.content)

Model IDs

• eleven_v3 - Eleven v3 (latest)
• eleven_multilingual_v2 - Multilingual v2
• eleven_turbo_v2_5 - Turbo v2.5
• eleven_flash_v2 - Flash v2

Rate Limits

Check your plan. Free tier has strict limits. Pro plans vary by subscription level.

---

Workflow Checklist

Before generating audio:

• Text normalized (numbers, dates, currencies spelled out if needed)
• Pronunciation checked for unusual names/terms
• Audio tags added for emotional beats (v3)
• Voice selected that matches content tone
• Stability setting appropriate for content type
• Text broken into reasonable chunks (long content)

After generating:

• Listen for pronunciation errors
• Check pacing feels natural
• Remove any spoken emotional directions if unwanted
• Trim silence at start/end
• Normalize audio levels if combining multiple clips

---

File Output Recommendations

For Web

• Format: MP3
• Bitrate: 128-192 kbps
• Sample rate: 44.1 kHz

For Podcast

• Format: MP3 or M4A
• Bitrate: 192-256 kbps
• Sample rate: 44.1 kHz
• Loudness: -16 LUFS (podcast standard)

For Video

• Format: WAV or high-bitrate MP3
• Sample rate: 48 kHz (matches video standard)
• Keep original quality for editing flexibility