ElevenLabs Speech Models

ElevenLabs is a leading AI audio company specializing in natural, expressive speech synthesis and accurate speech recognition. Through AvalAI, you can access 10 powerful ElevenLabs models for text-to-speech and speech-to-text tasks with industry-leading quality and performance.

Available Models

ElevenLabs models span two categories: Text-to-Speech (TTS) for generating natural-sounding audio from text, and Speech-to-Text (STT) for transcribing spoken audio with high accuracy.

Text-to-Speech Models

Eight TTS models are available via the v1/audio/speech endpoint, each optimized for different use cases:

Model	Description	Latency	Languages	Character Limit
eleven_v4	Latest and highest-quality, most emotive TTS	Higher	90+ languages	Not published
eleven_v4_turbo	Eleven v4 quality with low latency for real-time use	~150ms time to first speech	90+ languages	Not published
eleven_v3	Previous-generation emotionally rich multi-speaker TTS	Higher	70+ languages	5,000
eleven_multilingual_v2	Most lifelike, emotionally rich TTS	Higher	29 languages	10,000
eleven_turbo_v2_5	High quality, low-latency multilingual	~250-300ms	32 languages	40,000
eleven_turbo_v2	High quality, low-latency English	~250-300ms	English	40,000
eleven_flash_v2_5	Ultra-fast multilingual TTS	~75ms	32 languages	40,000
eleven_flash_v2	Ultra-fast English TTS	~75ms	English	40,000

Eleven v4 - Latest & Most Advanced

Eleven v4 (eleven_v4) is ElevenLabs' latest and most advanced text-to-speech model, and the first in a new generation of models built on an entirely new architecture. Compared with Eleven v3, it improves output quality, voice accuracy, consistency, emotion, delivery, audio tag handling, and language coverage. ElevenLabs describes it as a net upgrade over Eleven v3 in almost every case and recommends testing it with your own voices and content.

Eleven v4 interprets tone, pacing, emotion, character, and context from the text, so the same line can sound dramatic, tender, urgent, comedic, or conversational while keeping the speaker's identity. Because the model understands the context of a whole scene, multi-speaker dialogue sounds responsive rather than like separate lines stitched together.

Best for:

• Content creation and voiceovers where quality is the top priority
• Audiobook and long-form narration with consistent narrators and characters
• Character voice acting and dialogue for games and animation
• Dubbing and multilingual localization with a single voice

Key Features:

• Highest audio fidelity and emotional range in the ElevenLabs lineup
• More accurate speaker identity and consistency across generations, dialogue, narration, and regenerated lines
• Improved audio tags (for example [whispering], [shouting], [laughing], [excited, happy]) and natural-language delivery directions, including inline sound effects such as [light rain] or [phone buzzing]
• More reliable International Phonetic Alphabet (IPA) phoneme support for custom pronunciations
• More reliable request stitching when you chain generations for long-form content
• 90+ languages supported

Native accent handling: When the generated language matches the reference voice's language, the original accent is preserved. When the generated language is different, Eleven v4 produces fluent, natural-sounding speech in the target language instead of carrying over the reference voice's original accent. For example, a voice cloned from a Korean speaker produces natural English rather than Korean-accented English. This is a deliberate behavior change, not a bug. If your workflow depends on a voice keeping its native accent in other languages, test Eleven v4 before migrating. You can try audio tags to guide an accent, but results may vary.

Voice settings: Eleven v4 uses two voice settings: Stability, which controls how consistent delivery stays across generations, and Similarity, which controls how closely the output follows the reference voice. Professional Voice Clones also expose a Clone Strength setting in ElevenLabs. Style and Speed settings are not available, and SSML is not supported.

Voice cloning: Eleven v4 reproduces a source voice's timbre, cadence, delivery, accent, and loudness more faithfully than earlier models, including recording flaws. As a result, a voice can sound noticeably different from its Eleven v3 version. Compare both versions with your own content, and use clean source recordings.

Note

ElevenLabs is actively and continuously improving Eleven v4 after launch, so its behavior may change over time. Re-test your use case periodically.

Supported Languages: Afrikaans, Amharic, Arabic, Armenian, Assamese, Asturian, Azerbaijani, Belarusian, Bengali, Bosnian, Bulgarian, Burmese, Cantonese, Catalan, Cebuano, Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, Fula (Pulaar), Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Kannada, Kazakh, Korean, Kyrgyz, Lao, Lingala, Lithuanian, Luganda, Luxembourgish, Macedonian, Malay, Malayalam, Maltese, Mandarin Chinese, Māori, Marathi, Mongolian, Nepali, Norwegian Bokmål, Occitan, Odia, Pashto, Persian, Polish, Portuguese (Brazil), Punjabi, Romanian, Russian, Serbian, Shona, Sindhi, Slovak, Slovenian, Somali, Sorani Kurdish, Spanish (LatAm), Swahili, Swedish, Tajik, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, Wolof

Eleven v4 Turbo - Expressive Low Latency

Eleven v4 Turbo (eleven_v4_turbo) brings Eleven v4 technology to real-time use cases. It keeps very high quality and emotional delivery with a median inference latency of about 100ms and a median time to first speech of about 150ms, as measured by ElevenLabs over WebSocket streaming. It supports the same 90+ languages, native accent handling, and Stability and Similarity voice settings as Eleven v4.

Best for:

• Conversational voice agents that need expressive, natural speech
• Interactive voice experiences and games
• Customer support and healthcare agents that must pronounce specialized terms accurately
• Multilingual real-time applications served from a single voice

Eleven v3 - Previous Generation

Eleven v3 is ElevenLabs' previous-generation advanced speech synthesis model. It produces natural, life-like speech with high emotional range and contextual understanding across 70+ languages. For most new workloads, evaluate eleven_v4 first.

Best for:

• Character discussions with multiple interacting speakers
• Audiobook production with complex emotional delivery
• Emotional dialogue generation with high emotional range
• Natural multi-speaker dialogue support

Key Features:

• Natural multi-speaker dialogue support
• High emotional range and contextual understanding
• 70+ languages supported
• 5,000 character limit per request

Supported Languages: Afrikaans, Arabic, Armenian, Assamese, Azerbaijani, Belarusian, Bengali, Bosnian, Bulgarian, Catalan, Cebuano, Chichewa, Croatian, Czech, Danish, Dutch, English, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Kirghiz, Korean, Latvian, Lingala, Lithuanian, Luxembourgish, Macedonian, Malay, Malayalam, Mandarin Chinese, Marathi, Nepali, Norwegian, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Serbian, Sindhi, Slovak, Slovenian, Somali, Spanish, Swahili, Swedish, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese, Welsh

Eleven Multilingual v2 - Premium Quality

The most advanced, emotionally-aware speech synthesis model with rich emotional expression and consistent voice quality across languages.

Best for:

• Character voiceovers for gaming and animation
• Professional corporate videos and e-learning
• Multilingual projects requiring consistent voice
• Applications where lifelike speech is paramount

Supported Languages: English, Japanese, Chinese, German, Hindi, French, Korean, Portuguese, Italian, Spanish, Indonesian, Dutch, Turkish, Filipino, Polish, Swedish, Bulgarian, Romanian, Arabic, Czech, Greek, Finnish, Croatian, Malay, Slovak, Danish, Tamil, Ukrainian, Russian

Eleven Turbo v2.5 - Balanced Performance

High quality, low-latency model with excellent balance of quality and speed, supporting 32 languages.

Best for:

• Real-time voice agents and chatbots
• Interactive applications requiring quick response
• Multilingual production deployments
• Balance between quality and latency

Supported Languages: All Multilingual v2 languages plus Hungarian, Norwegian, Vietnamese

Eleven Flash v2.5 - Ultra-Low Latency

The fastest speech synthesis model designed for real-time applications with ~75ms latency.

Best for:

• Real-time voice agents (Agents Platform)
• Interactive games and applications
• Large-scale bulk TTS processing
• Cost-effective, fast speech synthesis

Speech-to-Text Models

Two state-of-the-art transcription models via the v1/audio/transcriptions endpoint:

Model	Description	Languages	Features
scribe_v2	State-of-the-art speech recognition	90+ languages	Diarization, timestamps, entity detection
scribe_v1	Legacy transcription model	90+ languages	Basic transcription

Scribe v2 - Advanced Transcription

State-of-the-art speech recognition with comprehensive features:

• Accurate Transcription: 90+ languages supported
• Keyterm Prompting: Up to 100 terms for domain-specific accuracy
• Entity Detection: Up to 56 entity types recognized
• Word-level Timestamps: Precise timing for each word
• Speaker Diarization: Identify up to 32 speakers
• Dynamic Audio Tagging: Automatic audio event detection
• Smart Language Detection: Automatic language identification

Pricing

Text-to-Speech Pricing

ElevenLabs TTS models are priced per second of generated audio:

Model	Output Cost (per second)	Best For
eleven_v4	$0.0014/sec	Latest, highest quality, most emotive
eleven_v4_turbo	$0.0007/sec	Expressive, low latency, real-time
eleven_v3	$0.005/sec	Previous generation, multi-speaker
eleven_turbo_v2	$0.0025/sec	English, low latency
eleven_turbo_v2_5	$0.0025/sec	Multilingual, balanced
eleven_flash_v2	$0.0025/sec	English, fastest
eleven_flash_v2_5	$0.0025/sec	Multilingual, fastest
eleven_multilingual_v2	$0.005/sec	Highest quality

Speech-to-Text Pricing

Scribe models are priced per second of input audio:

Model	Input Cost (per second)
scribe_v2	$0.00009722/sec
scribe_v1	$0.00009722/sec

API Endpoints

ElevenLabs models are available through AvalAI's OpenAI-compatible API:

• Text-to-Speech: v1/audio/speech
• Transcriptions: v1/audio/transcriptions

Example Usage

Text-to-Speech Generation

from openai import OpenAI
from pathlib import Path

client = OpenAI(api_key="your-avalai-api-key", base_url="https://api.avalai.ir/v1")

# Generate speech with Flash model for low latency
speech_file = Path("./speech.mp3")

response = client.audio.speech.create(
    model="eleven_flash_v2_5",
    input="Hello! This is a demonstration of ElevenLabs text-to-speech via AvalAI.",
    voice="coral",
)

response.stream_to_file(speech_file)
print(f"Audio saved to {speech_file}")

cURL Example - TTS

curl https://api.avalai.ir/v1/audio/speech \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "eleven_flash_v2_5",
    "input": "The quick brown fox jumped over the lazy dog.",
    "voice": "coral"
  }' \
  --output speech.mp3

Expressive TTS with Eleven v4

Use eleven_v4 for the highest quality and audio tags to direct delivery. Do not send speed with Eleven v4 models, because Speed is not available for them.

# Use Eleven v4 for expressive, high-quality narration
response = client.audio.speech.create(
    model="eleven_v4",
    input="[excited, happy] Big news, everyone. Our summer menu drops this Friday.",
    voice="alloy",
)

response.stream_to_file("eleven_v4_speech.mp3")

For low-latency voice agents, change only model to eleven_v4_turbo.

High Quality TTS with Multilingual v2

# Use Multilingual v2 for highest quality output
response = client.audio.speech.create(
    model="eleven_multilingual_v2",
    input="Welcome to our platform. We're delighted to have you here.",
    voice="alloy",
)

response.stream_to_file("premium_speech.mp3")

Speech Transcription

curl https://api.avalai.ir/v1/audio/transcriptions \
  -H "Authorization: Bearer $AVALAI_API_KEY" \
  -F file=@audio.mp3 \
  -F model=scribe_v2

Example Response:

{
  "text": "Hello, this is a sample transcription from the audio file.",
  "task": "transcribe",
  "language": "en",
  "duration": 3.5,
  "words": [
    {
      "word": "Hello",
      "start": 0.0,
      "end": 0.35
    },
    {
      "word": "this",
      "start": 0.4,
      "end": 0.55
    },
    {
      "word": "is",
      "start": 0.55,
      "end": 0.65
    },
    {
      "word": "a",
      "start": 0.65,
      "end": 0.72
    },
    {
      "word": "sample",
      "start": 0.72,
      "end": 1.0
    }
  ]
}

Available Voices

ElevenLabs models support multiple voices. Common voices include:

• alloy - Neutral, professional
• coral - Warm, friendly
• echo - Clear, expressive
• fable - Storytelling, narrative
• nova - Energetic, upbeat
• onyx - Deep, authoritative
• sage - Calm, wise
• shimmer - Bright, cheerful

Model Selection Guide

Use Case	Recommended Model
Expressive real-time voice agents	eleven_v4_turbo
Lowest-latency voice agents	eleven_flash_v2_5
Interactive applications	eleven_v4_turbo, eleven_flash_v2_5, or eleven_turbo_v2_5
Audiobook production	eleven_v4
Character voiceovers	eleven_v4
Dubbing and localization	eleven_v4
Bulk processing	eleven_flash_v2_5
English-only projects	eleven_turbo_v2 or eleven_flash_v2
Multilingual content	eleven_v4, eleven_v4_turbo, or eleven_multilingual_v2
High-accuracy transcription	scribe_v2

Best Practices

For Text-to-Speech

1. Choose the right model: Use Eleven v4 for the highest quality, Eleven v4 Turbo for expressive real-time speech, and Flash for the lowest latency
2. Voice selection: Match voice characteristics to your content
3. Text preparation: Clean and format text for natural speech patterns
4. Language consistency: Use language-appropriate models for best results
5. Direct delivery with audio tags: With Eleven v4 models, use tags such as [whispering] or [laughing] instead of SSML, which is not supported

For Transcription

1. Audio quality: Higher quality input produces better transcriptions
2. Use keyterm prompting: Add domain-specific terms for accuracy
3. Speaker diarization: Enable when multiple speakers are present
4. Language hints: Specify language when known for better accuracy

Related Resources

• Audio API Reference
• Official ElevenLabs API Documentation
• ElevenLabs Models Overview
• Pricing