ElevenLabs Cheat Sheet
Current quick reference for the ElevenLabs platform in 2026
This cheat sheet covers the current ElevenLabs stack: text-to-speech (TTS), model selection, voice library and cloning, API usage, streaming, language support, and where agents and speech-to-text fit into the platform.
Reading time:≈ 7 minutes
Last updated:March 30, 2026
Current ElevenLabs Platform
ElevenLabs is no longer just a text-to-speech tool. The platform now spans TTS, speech-to-text, voices, agents, and other audio workflows.
Text to Speech
Core TTS models include eleven_flash_v2_5, eleven_multilingual_v2, and eleven_v3.
Speech to Text
Scribe v2 and realtime STT support transcription, timestamps, captions, and meeting-style audio workflows.
Voices
Use Voice Library, Instant Voice Cloning, Professional Voice Cloning, and Voice Design depending on speed, quality, and sharing needs.
Agents
ElevenLabs now supports agent-style, low-latency voice experiences where TTS and realtime audio matter more than one-off generation.
Music & Sound
The platform also covers music and sound effects, so ElevenLabs now sits closer to a broader audio generation stack than a TTS-only API.
Dubbing & Cleanup
Dubbing, voice isolation, and related audio cleanup workflows are now part of the product surface many users expect from ElevenLabs.
What changed since older cheat sheets?
- Older v1-era references are outdated.
- Turbo-first recommendations are no longer the best default framing.
- The product is broader now: TTS, STT, voices, agents, and audio workflows.
Quick Start Guide
Get results in 5 minutes
1
Complete Working Example
Copy this cURL command, replace API_KEY and run it to generate your first voice:
curl -X POST https://api.elevenlabs.io/v1/text-to-speech/21m00Tcm4TlvDq8ikWAM \ -H "xi-api-key: YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "text": "Welcome to ElevenLabs. This is your first generated voice.", "model_id": "eleven_flash_v2_5", "voice_settings": { "stability": 0.5, "similarity_boost": 0.75 } }' \ --output speech.mp3
Voice ID used:
Get your API key: elevenlabs.io → Developer → API Key
21m00Tcm4TlvDq8ikWAM (Rachel - default female voice) Get your API key: elevenlabs.io → Developer → API Key
2
Choose Your Model
Need emotion & drama?
→
eleven_v3 Dialogue, characters, storytelling
Need speed (real-time)?
→
eleven_flash_v2_5 Chatbots, live agents (75ms latency)
Need quality & languages?
→
eleven_multilingual_v229 languages, premium quality
Need a safe default?
→
eleven_flash_v2_5 or eleven_multilingual_v2 Flash for speed, Multilingual for polished narration
3
Tune Voice Settings
Stability
0.0 - 1.0Low (0.3): Varied & expressive
High (0.8): Consistent & predictable
High (0.8): Consistent & predictable
Similarity Boost
0.0 - 1.0 Higher = Closer to original voice
Sweet spot: 0.75 for most cases
Sweet spot: 0.75 for most cases
Style Exaggeration
0.0 - 1.0 Amplifies voice's natural style
Start at 0.0, increase if needed
Start at 0.0, increase if needed
Models Comparison
Choose the right model for your use case
| Model | Best For | Latency | Languages | Char Limit | Key Features |
|---|---|---|---|---|---|
Flash v2.5 eleven_flash_v2_5 | Real-time applications Chatbots, live agents, conversational AI | 75ms Ultra-fast | 32 languages | 40,000 ~40 min audio | • Lowest latency • Optimized for speed • WebSocket streaming • Good quality/speed balance |
Multilingual v2 eleven_multilingual_v2 | Premium quality Audiobooks, podcasts, videos | ~1-2s Slower | 29 languages Premium quality | 10,000 Longer-form narration | • Highest quality • Best for long-form content • Natural prosody • Accent preservation |
Eleven v3 eleven_v3 | Emotional & expressive Storytelling, character voices, drama | ~1-2s Slower | 70+ languages Most supported | 5,000 Designed for expressive output | • Multi-speaker dialogue • Audio tags [laughs] [whispers] • Most emotional depth • Character acting |
Scribe v2 scribe_v2 | Speech-to-text Transcription, timestamps, diarization | ~150ms realtime Realtime variant available | 90+ languages Most coverage | Audio file based | • Converts speech to text • Word-level timestamps • Speaker diarization • High accuracy |
Need Speed?
Real-time, conversational AI, low latency required
→ Flash v2.5
Need Quality?
Audiobooks, premium multilingual content, natural prosody
→ Multilingual v2
Need Emotion?
Drama, character voices, expressive storytelling
→ Eleven v3
Need STT?
Transcription, realtime captions, meeting audio
→ Scribe v2 / Realtime
Text Control & Formatting
Control timing, pronunciation, and speech patterns
Pauses & Timing
Break Tags (up to 3 seconds)
<break time="1.5s" />Insert natural pauses between sentences
Dots for Endpoints
Sentence. <break time="2s" /> Next.Combine with dots for cleaner breaks
Warning: Too many breaks = instability
Pronunciation Control
CMU Arpabet (Recommended)
<phoneme alphabet="cmu-arpabet" ph="M AE1 D IH0 S AH0 N">Madison</phoneme>IPA (International)
<phoneme alphabet="ipa" ph="ˈæktʃuəli">actually</phoneme>Compatible: Flash v2, Turbo v2, English v1
Alias Tags
For Other Models (v3, Multilingual)
<lexeme> <grapheme>UN</grapheme> <alias>United Nations</alias> </lexeme>Case-sensitive
Exact matching
Context-aware
Understands usage
Speed & Pacing
Speed Range
0.7 - 1.2Slow (0.7)Deliberate narration
Normal (1.0)Default pace
Fast (1.2)Quick delivery
Use narrative style in text for natural pacing instead of extreme speeds
Pronunciation Dictionaries (.PLS)
What They Are
- Upload custom pronunciation files
- Reusable across multiple requests
- Combine phonemes + aliases
- Case-sensitive matching
Useful Tools
Sequitur G2P
Grapheme-to-phoneme conversion
Phonetisaurus
Phoneme sequence modeling
eSpeak
Text-to-phoneme synthesis
Emotion & Expression
Make AI voices sound more human and expressive without overfitting to fragile prompt tricks
Narrative Context
Add emotional context around dialogue for natural expression
Sadness Example
"You're leaving?" she asked, her voice trembling with sadness. Triumph Example
"That's it!" he exclaimed triumphantly. Works with all models - most reliable method for emotion
v3 Audio Tags
Special tags for Eleven v3 model only
Voice Tags
[laughs][whispers][sighs][exhales][sarcastic][curious][excited][crying] Expressive Tags
[applause][clapping][gasps]Special Tags & Effects
Accent Control
[strong French accent] Bonjour! Useful for experimentation, but validate results per voice and model.
Other Special Tags
[sings]Singing voicePunctuation Techniques
CAPITALS for Emphasis
This is AMAZING!Adds stress and volume
Ellipses for Weight
I don't know... maybe...Creates hesitation and pauses
Question Marks
You did what?Natural upward inflection
Eleven v3 Mode Settings
Choose the right mode for your emotional needs
Creative
• Most emotional
• Prone to hallucinations
• Best for dramatic content
Use for: Character acting, storytelling
Natural
• Balanced approach
• Closest to original voice
• Good responsiveness
Use for: General purpose, audiobooks
Robust
• Very stable output
• Less responsive to prompts
• Consistent quality
Use for: Technical content, narration
Voice Selection & Settings
Choose the right voice source, clone type, and settings for your use case
Current ElevenLabs Voice Stack
Voice Library
Browse marketplace and prebuilt voices for fast testing.
IVC
Instant Voice Cloning for fast turnaround from a small sample.
PVC
Professional Voice Cloning for the highest quality and shareable clones.
Voice Design
Create a new synthetic voice instead of cloning an existing speaker.
Voice Library
Pre-built Collections
Browse prebuilt and community voices in the Voice Library marketplace
Emotionally diverse range
Professional quality
Ready to use
Popular Voices
Rachel - Neutral female Adam - Deep male Bella - Soft female Tip: use the library for rapid testing before investing in a custom clone.
Voice Cloning Types
IVC (Instant Voice Cloning)
Upload 1-2 minutes of audio, instant results
Best for quick prototyping and internal testing
PVC (Professional Voice Cloning)
3+ hours of audio, highest quality
Studio-grade quality
Longer processing time
Shareability
Professional Voice Clones can be shared in the Voice Library; treat IVC as a faster, lighter path.
Voice Settings Presets
Narration
Stability
0.6-0.8Similarity
0.7Style
0.2-0.4 Audiobooks, documentation, training content
Customer Support
Stability
0.7-0.9Similarity
0.9Style
0.0-0.2 Clear, consistent voice for customer communication
Characters
Stability
0.3-0.5Similarity
0.8Style
0.7-1.0 Dramatic, highly expressive character voices
Speaker Boost
What It Does
Enhances voice clarity and presence, reduces artifacts
When to Use
Professional recordings
Audiobooks & podcasts
Not for real-time (adds latency)
Selection Tips
Test Multiple Voices
Same text sounds different across voices
Match Content Type
Formal content needs neutral voices
Consider Audience
Age, region, and preferences matter
Regenerate Feature
Generate a few takes before locking a voice for production
WebSocket Streaming
Real-time audio generation patterns for low-latency ElevenLabs workflows
Connection Setup
WebSocket Endpoint
wss://api.elevenlabs.io/v1/text-to-speech/{voice_id}/stream-input Required Headers
xi-api-key: YOUR_API_KEY Use Flash v2.5 for best streaming performance. Confirm endpoint parameters against current docs before production rollout.
Message Protocol
Send Configuration (JSON)
{ "text": "Hello world", "model_id": "eleven_flash_v2_5", "voice_settings": { "stability": 0.5, "similarity_boost": 0.75 } } Receive Audio Chunks
Audio chunks stream back incrementally; exact framing and formats can vary by integration.
Chunk Management
Send Text in Chunks
ws.send(JSON.stringify({ text: "First chunk. " })) ws.send(JSON.stringify({ text: "Second chunk." })) Stream text incrementally for faster response, ideally at phrase or sentence boundaries.
Chunking Guidance
Smaller chunksLower latency
Sentence-sized chunksBest balance
Very large chunksHigher latency
Flushing & Completion
Flush Message
ws.send(JSON.stringify({ flush: True })) Flush any remaining buffered text for a clean ending.
Completion Signal
Watch for the service completion event defined in the current WebSocket docs.
Buffering Strategy
Initial Buffer
Wait ~250ms before playback to avoid stutter
Rolling Buffer
Always keep 150-200ms of audio ahead
Backpressure
Pause sending text if playback lags
Error Handling
Common Errors
- 401 - Invalid API key
- 429 - Rate limited
- 400 - Invalid chunk format
Retries
Exponential backoff (250ms → 4s)
Performance Tips
Minimize Latency
- Use Flash v2.5 model
- Send 200-500 char chunks
- Flush at sentence breaks
- Keep connection alive
Optimize Quality
- Buffer 300-500ms initially
- Use complete sentences
- Avoid mid-word chunks
- Proper punctuation
Handle Errors
- Implement reconnection
- Monitor connection health
- Handle rate limits
- Log failed chunks
API Quick Reference
Essential endpoints, model IDs, and current usage patterns
Text-to-Speech (POST)
Endpoint
POSThttps://api.elevenlabs.io/v1/text-to-speech/{voice_id} Required Headers
xi-api-key: YOUR_API_KEYContent-Type: application/json Request Body
{ "text": "Your text here", "model_id": "eleven_flash_v2_5", "voice_settings": { "stability": 0.5, "similarity_boost": 0.75, "style": 0.0, "use_speaker_boost": true } }Python Example
import requests url = f"https://api.elevenlabs.io/v1/text-to-speech/{voice_id}" headers = { "xi-api-key": API_KEY, "Content-Type": "application/json" } data = { "text": "Hello world", "model_id": "eleven_flash_v2_5" } response = requests.post(url, json=data, headers=headers) with open("output.mp3", "wb") as f: f.write(response.content)
JavaScript Example
const response = await fetch(url, { method: 'POST', headers: { 'xi-api-key': API_KEY, 'Content-Type': 'application/json' }, body: JSON.stringify({ text: 'Hello world', model_id: 'eleven_flash_v2_5' }) }); const audioBuffer = await response.arrayBuffer(); const blob = new Blob([audioBuffer], { type: 'audio/mpeg' }); const urlObject = URL.createObjectURL(blob);
List Voices (GET)
GEThttps://api.elevenlabs.io/v1/voices Returns all available voices with metadata
List Models (GET)
GEThttps://api.elevenlabs.io/v1/models See available model IDs and capabilities
Speech-to-Text (POST)
POSThttps://api.elevenlabs.io/v1/speech-to-text Use Scribe v2 for transcription, timestamps, and diarization workflows.
Voice Cloning (POST)
POSThttps://api.elevenlabs.io/v1/voices/add- Upload audio files
- Provide descriptive name
- Choose cloning type (IVC/PVC)
- Optional: description, labels
Current Model IDs to Know
Fast TTS
eleven_flash_v2_5 Premium TTS
eleven_multilingual_v2 Expressive TTS
eleven_v3Language Support
Current language coverage depends on model, with Eleven v3 offering the broadest TTS support
Language Coverage by Model
Flash v2.5
32
Major global languages
Multilingual v2
29
Premium quality support
Eleven v3
70+
Most comprehensive
Language Codes
{ "text": "Hola, ¿cómo estás?", "model_id": "eleven_multilingual_v2", "language_code": "es" } Set
language_code when the endpoint and model support it for more predictable output. Multilingual Tips
Consistency
Keep punctuation style consistent with the language
Context
Provide regional context (e.g., Mexico vs Spain)
Accents
Prefer voice choice and clean localized text over heavy prompt tricks
Regional Variants
- Spanish: es-ES, es-MX, es-AR
- Portuguese: pt-PT, pt-BR
- English: en-US, en-GB, en-AU
- French: fr-FR, fr-CA
Code Switching
"Welcome to Mexico City. Bienvenidos a la Ciudad de México." Use full sentences per language for best effect
Model-Specific Notes
Flash v2.5
- Best for realtime TTS
- 32 languages
- Strong default for agents
- Use for low-latency playback
Multilingual v2
- Best for polished multilingual narration
- 29 languages
- Better long-form balance
- Use when quality matters more than speed
Eleven v3
- Broadest TTS language coverage
- 70+ languages
- Best expressive model
- Use for dialogue and emotional delivery
Special Characters
Accented Characters
á é í ó ú ñ ç ü ö ä
Cyrillic Script
Russian, Ukrainian, Bulgarian
Right-to-Left
Arabic, Hebrew support
Diacritics
Proper pronunciation handling
Automatic Detection
How It Works
Some workflows can infer language automatically, but explicit configuration is safer.
Works best for common, clean inputs
Less accurate than explicit code
Always prefer language_code
Troubleshooting
Common issues and solutions
Audio Quality Issues
Robotic/Unnatural Sound
- Lower stability (try 0.3-0.5)
- Increase similarity boost
- Add narrative context
- Try different voice
Inconsistent Output
- Increase stability (0.7-0.8)
- Switch from v3 to Flash or Multilingual for more stable output
- Remove excessive breaks
- Simplify text formatting
Artifacts/Glitches
- Enable speaker boost
- Reduce style exaggeration
- Fix text punctuation
- Regenerate (use 3x feature)
Voice Cloning Problems
Poor Voice Match
- Use 2+ minutes of audio
- Ensure clean recording
- Single speaker only
- No background noise
Clone Sounds Different
- Match audio style to use case
- Adjust similarity boost
- Try PVC instead of IVC
- Test clone quality across current TTS models
Upload Fails
- Check file format (MP3/WAV)
- Max 10 files per upload
- 1-2 min total duration
- Remove silence/pauses
Pronunciation Issues
Mispronounced Names
- Use phoneme tags
- Add pronunciation dictionary
- Provide phonetic hints
- Spell out syllables
Foreign Words
- Set language_code
- Use narrative context
- Try accent tags
- Break into syllables
Latency Issues
Slow Responses
- Use Flash v2.5
- Shorten text chunks
- Avoid complex tags
- Use WebSocket streaming
Queue Delays
- Retry after delay
- Check status page
- Check plan limits and current service guidance
- Batch non-urgent jobs
Emotion Not Working
Flat Delivery
- Add contextual cues
- Use narrative descriptions
- Lower stability
- Use v3 tags
Overacting
- Increase stability
- Reduce style exaggeration
- Use calmer language
- Switch to Flash or Multilingual for more controlled output
Character Limit Issues
Text Too Long
- Flash v2.5: 40k limit
- Multilingual v2: 10k limit
- Eleven v3: 5k limit
- Split by sentences
Streaming Overflows
- Send 200-500 char chunks
- Flush after sections
- Use queue for long form
- Store progress markers
ElevenLabs FAQ
Fast answers to common ElevenLabs questions people search before using the API or picking a model.
What is the best ElevenLabs model right now?
It depends on the job: use Flash v2.5 for low latency, Multilingual v2 for polished multilingual narration, and Eleven v3 for expressive dialogue or character-style output.
Is ElevenLabs only for text to speech?
No. ElevenLabs now spans text-to-speech, speech-to-text, voice cloning, voice design, streaming, agents, dubbing, and other audio workflows.
Which ElevenLabs model should I use for real-time apps?
Start with eleven_flash_v2_5 for low-latency generation, especially when building chat, assistant, or agent-style voice interfaces.
What is the difference between IVC and PVC?
Instant Voice Cloning is faster and lighter for quick tests, while Professional Voice Cloning is the higher-quality path for serious production workflows and shareable cloned voices.
Does ElevenLabs support multiple languages?
Yes. Coverage varies by model, with Eleven v3 offering the broadest TTS language support, while Multilingual v2 remains a strong choice for high-quality multilingual narration.
More Cheatsheets
Other quick-reference guides you might find useful.
