elevenlabs-tts

Install the belt CLI skill: npx skills add belt-sh/cli

ElevenLabs Text-to-Speech

Premium text-to-speech with 22+ voices via inference.sh CLI. Each ElevenLabs model is its own app, at one price.

Quick Start

Requires the inference.sh CLI (belt). Install instructions

belt login

# Generate speech with Eleven v4
belt app run elevenlabs/eleven-v4 --input '{"text": "Hello, welcome to our product demo.", "voice": "aria"}'

The result has an audio URL to the generated file.

Which App

App	Model	Use it for	Text limit	Price per 1K characters
elevenlabs/eleven-v4	Eleven v4	Newest and most expressive; 90+ languages; audio tags	10,000	$0.08
elevenlabs/eleven-v4-turbo	Eleven v4 Turbo	The low-latency v4; realtime voice agents	10,000	$0.04
elevenlabs/eleven-v3	Eleven v3	Previous generation; 70+ languages; audio tags	5,000	$0.08
elevenlabs/eleven-multilingual-v2	Eleven Multilingual v2	Stable long-form narration	10,000	$0.08
elevenlabs/eleven-flash-v2-5	Eleven Flash v2.5	Fastest and cheapest; long texts and high volume	40,000	$0.04

• Every app has run: the whole text in, an audio file out.
• The two v4 apps also have realtime: text streamed in, speech streamed back as it is written (see Realtime Speech).
• elevenlabs/eleven-v4 and elevenlabs/eleven-v3 also have dialogue: several voices in one take (see Dialogue).
• Check the current price with belt app pricing elevenlabs/eleven-v4.

elevenlabs/tts is the older app that takes the model as an input. It still works; prefer the per-model apps above.

Voice Library

Female Voices

Voice	Style
aria	American, conversational
alice	British, confident
bella	American, warm
jessica	American, expressive
laura	American, professional
lily	British, soft
sarah	American, friendly

Male Voices

Voice	Style
george	British, authoritative
adam	American, deep
bill	American, mature
brian	American, conversational
callum	Transatlantic, intense
charlie	Australian, natural
chris	American, casual
daniel	British, commanding
eric	American, friendly
harry	American, young
liam	American, articulate
matilda	American, warm
river	American, confident
roger	American, authoritative
will	American, bright

Examples

Basic Speech

belt app run elevenlabs/eleven-v4 --input '{"text": "Welcome to our quarterly earnings presentation.", "voice": "george"}'

Audio Tags (v4 and v3)

Tags in square brackets are performed, not read aloud. They are free-form directions: emotion, delivery, accents, sounds.

belt app run elevenlabs/eleven-v4 --input '{
  "text": "[excited] We did it! [whispers] Nobody saw this coming. [laughs] What a day.",
  "voice": "jessica"
}'

Common tags: [laughs], [whispers], [sighs], [excited], [sad], [shouts], [slow], [fast]. The other models read tags aloud, so leave them out there.

Choose a Model

# Long narration, stable delivery
belt app run elevenlabs/eleven-multilingual-v2 --input '{
  "text": "Chapter one. It was a quiet morning in the harbour town.",
  "voice": "daniel"
}'

# High volume at the lowest price
belt app run elevenlabs/eleven-flash-v2-5 --input '{
  "text": "Your order has shipped and will arrive on Thursday.",
  "voice": "sarah"
}'

Voice Tuning

belt app run elevenlabs/eleven-v3 --input '{
  "text": "Fine-tune the voice characteristics for your use case.",
  "voice": "bella",
  "stability": 0.3,
  "similarity_boost": 0.9,
  "style": 0.4
}'

Parameter	Range	Effect	Apps
stability	0-1	Higher = more consistent, lower = more expressive	all
similarity_boost	0-1	Higher = closer to original voice character	all
style	0-1	Higher = more style exaggeration	v3, Multilingual v2, Flash v2.5
use_speaker_boost	true/false	Enhances speaker clarity	v3, Multilingual v2, Flash v2.5

Eleven v4 and v4 Turbo have two settings only, stability and similarity_boost. Shape their delivery with audio tags.

Custom Voices

Pass voice_id to use a cloned or designed voice (from elevenlabs/voice-clone or elevenlabs/voice-design). It overrides voice.

belt app run elevenlabs/eleven-v4 --input '{"text": "This is my own voice.", "voice_id": "YOUR_VOICE_ID"}'

Output Formats

# High-quality MP3
belt app run elevenlabs/eleven-v4 --input '{
  "text": "High quality audio output.",
  "voice": "daniel",
  "output_format": "mp3_44100_192"
}'

Format	Description
mp3_44100_128	MP3 at 44.1kHz, 128kbps (default)
mp3_44100_192	MP3 at 44.1kHz, 192kbps
pcm_16000	Raw PCM at 16kHz
pcm_22050	Raw PCM at 22.05kHz
pcm_24000	Raw PCM at 24kHz
pcm_44100	Raw PCM at 44.1kHz

Other Languages

Write the text in the language you want spoken.

# Spanish
belt app run elevenlabs/eleven-v4 --input '{
  "text": "Hola, bienvenidos a nuestra presentación.",
  "voice": "aria"
}'

# French
belt app run elevenlabs/eleven-v4 --input '{
  "text": "Bonjour, bienvenue à notre démonstration.",
  "voice": "alice"
}'

Dialogue

dialogue speaks a conversation in one take and returns one file. Each line has its own voice.

belt app run elevenlabs/eleven-v4 --function dialogue --input '{
  "segments": [
    {"text": "[excited] Did you hear? The release went out last night.", "voice": "jessica"},
    {"text": "[surprised] Already? I thought that was next week.", "voice": "george"},
    {"text": "[laughs] Nope. Everything shipped.", "voice": "jessica"}
  ]
}'

Up to 10 different voices. Keep one dialogue under about 2,000 characters and split a longer script into scenes. The price is the app's rate per character, counted across all lines.

Realtime Speech

realtime on elevenlabs/eleven-v4-turbo and elevenlabs/eleven-v4 is a live function: you stream text as it is written (an LLM's tokens, say) and get speech back while it is still coming. It runs over a socket, so call it from the SDK or the app's page, not with belt app run.

# pip install "inferencesh[async]"
import asyncio
from inferencesh import async_inference

async def main():
    client = async_inference(api_key="...")
    task, session = await client.live({
        "app": "elevenlabs/eleven-v4-turbo",
        "function": "realtime",
        "input": {"voice": "george"},
    })

    async def speak():
        for piece in ["Hello there. ", "This is spoken ", "while it is being written. "]:
            await session.send({"events": {"type": "text", "text": piece}})
        await session.send({"events": {"type": "end"}})   # finish speaking, then end

    asyncio.create_task(speak())
    async for frame in session:
        if isinstance(frame, bytes):
            play(frame)            # raw PCM, 16-bit mono, 24 kHz
        else:
            print(frame)           # {"characters": 54, "seconds": 3.9}

asyncio.run(main())

What you send, as {"events": {...}}:

Event	Meaning
{"type": "text", "text": "..."}	Text to speak. Send it in pieces as it arrives.
{"type": "flush"}	Speak what was sent so far without waiting for more
{"type": "end"}	No more text: finish speaking, then end the session

• Speech starts once about 40 characters and 8 words have arrived. For a short line, send flush.
• Closing the socket ends the session at once. Send end when you want to hear the rest first.
• The voice (voice or voice_id), stability and language_code are set when the session opens.
• A session ends by itself after idle_minutes (default 2) without text.
• You pay for the characters sent, at the app's rate.

Voice + Video Workflow

# 1. Generate voiceover
belt app run elevenlabs/eleven-v4 --input '{
  "text": "Introducing the future of AI-powered content creation.",
  "voice": "george"
}' > voiceover.json

# 2. Create talking head video
belt app run bytedance/omnihuman-1-5 --input '{
  "image": "https://portrait.jpg",
  "audio": "<audio-url-from-step-1>"
}'

Use Cases

• Voiceovers: Product demos, explainer videos, commercials
• Audiobooks: Long-form narration with consistent voices
• Podcasts: AI hosts with natural delivery
• E-learning: Course narration in multiple languages
• Accessibility: High-quality screen reader content
• IVR: Professional phone system messages
• Voice agents: Speak an LLM's answer while it is still being written, with realtime
• Video Narration: Documentary and social media content

Related Skills

# ElevenLabs multi-speaker dialogue
npx skills add inference-sh/skills@elevenlabs-dialogue

# ElevenLabs voice changer
npx skills add inference-sh/skills@elevenlabs-voice-changer

# ElevenLabs sound effects
npx skills add inference-sh/skills@elevenlabs-sound-effects

# All TTS models (Kokoro, DIA, Chatterbox, Inworld TTS, and more)
npx skills add inference-sh/skills@text-to-speech

# Full platform skill (all apps)
npx skills add inference-sh/skills@infsh-cli

Browse all audio apps: belt app list --category audio