User asks to generate speech, narrate text, create a voice-over, compose music, or produce a sound effect.
VideoAgent Audio Studio is a smart audio dispatcher. It analyzes your request and routes it to the best available model — ElevenLabs for speech and music, fal.ai for fast SFX — and returns a ready-to-use audio URL.
Quick Reference
Request Type
Best Model
Latency
Narrate text / Voice-over
elevenlabs-tts-v3
~3s
Low-latency TTS (real-time)
elevenlabs-tts-turbo
<1s
Background music
cassetteai-music
~15s
Sound effect
elevenlabs-sfx
~5s
Clone a voice from audio
elevenlabs-voice-clone
~10s
How to Use
1. Start the AudioMind server (once per session)
bash
{
baseDir
}
/tools/start_server.sh
This starts the ElevenLabs MCP server on port 8124. The skill uses it for all audio generation.
2. Route the request
Analyze the user's request and call the appropriate tool via the MCP server:
Text-to-Speech (TTS)
When user asks to "narrate", "read aloud", "say", or "create a voice-over":
model_id: "eleven_multilingual_v2" # Use "eleven_turbo_v2_5" for low latency
Music Generation
When user asks to "compose", "create background music", or "make a soundtrack":
Use MCP tool: text_to_sound_effects (via cassetteai-music on fal.ai)
prompt: ""
duration_seconds:
Sound Effect (SFX)
When user asks for a specific sound (e.g., "a door creaking", "rain on a window"):
Use MCP tool: text_to_sound_effects
text: ""
duration_seconds: <1-22>
Voice Cloning
When user provides an audio sample and wants to clone the voice:
Use MCP tool: voice_add
name: ""
files: [""]
Example Conversations
User:
"Voice this text for me: Welcome to our product launch"
→ Route to: text_to_speech
text: "Welcome to our product launch"
voice_id: "JBFqnCBsd6RMkjVDRZzb"
model_id: "eleven_multilingual_v2"
🎙️ Voiceover done!
Listen here
User:
"Generate 60 seconds of relaxing background music for a podcast"
→ Route to: cassetteai-music (fal.ai)
prompt: "relaxing lo-fi background music for a podcast, gentle piano and soft beats, 60 seconds"
duration_seconds: 60
🎵 Background music ready!
Listen here
User:
"Generate a sci-fi style door opening sound effect"
→ Route to: text_to_sound_effects
text: "a futuristic sci-fi door sliding open with a hydraulic hiss"
duration_seconds: 3
Setup
Required
Set
ELEVENLABS_API_KEY
in
~/.openclaw/openclaw.json
:
{
"skills"
:
{
"entries"
:
{
"videoagent-audio-studio"
:
{
"enabled"
:
true
,
"env"
:
{
"ELEVENLABS_API_KEY"
:
"your_elevenlabs_key_here"
}
}
}
}
}
Get your key at
elevenlabs.io/app/settings/api-keys
.
Optional (for fal.ai music & SFX models)
"FAL_KEY"
:
"your_fal_key_here"
Get your key at
fal.ai/dashboard/keys
.
Self-Hosting the Proxy
The
cli.js
connects to a hosted proxy by default. If you want full control — or need to serve users in regions where
vercel.app
is blocked — you can deploy your own instance from the
proxy/
directory.
Quick Deploy (Vercel)
cd
proxy
npm
install
vercel
--prod
Environment Variables
Set these in your Vercel project (Dashboard → Settings → Environment Variables):
Variable
Required For
Where to Get
ELEVENLABS_API_KEY
TTS, SFX, Voice Clone
elevenlabs.io/app/settings/api-keys
FAL_KEY
Music generation
fal.ai/dashboard/keys
VALID_PRO_KEYS
(Optional) Restrict access
Comma-separated list of allowed client keys
Point cli.js to Your Proxy
export
AUDIOMIND_PROXY_URL
=
"https://your-domain.com/api/audio"
Or set it in
~/.openclaw/openclaw.json
:
{
"skills"
:
{
"entries"
:
{
"videoagent-audio-studio"
:
{
"env"
:
{
"AUDIOMIND_PROXY_URL"
:
"https://your-domain.com/api/audio"
}
}
}
}
}
Custom Domain (Recommended)
If your users are in mainland China, bind a custom domain in Vercel Dashboard → Settings → Domains to avoid DNS issues with
vercel.app
.
Model Reference
Model ID
Type
Provider
Notes
eleven_multilingual_v2
TTS
ElevenLabs
Best quality, supports 29 languages
eleven_turbo_v2_5
TTS
ElevenLabs
Ultra-low latency, ideal for real-time
eleven_monolingual_v1
TTS
ElevenLabs
English only, fastest
cassetteai-music
Music
fal.ai
Reliable, fast music generation
elevenlabs-sfx
SFX
ElevenLabs
High-quality sound effects (up to 22s)
elevenlabs-voice-clone
Clone
ElevenLabs
Clone any voice from a short audio sample
Changelog
v3.0.0
Simplified routing table
Removed unstable/offline models from the main reference. The skill now only surfaces models that reliably work.
Clearer use-case triggers
Added "Use when" section so the agent activates this skill at the right moment.
Unified setup
Single
ELEVENLABS_API_KEY
is all you need to get started.
FAL_KEY
is now optional.
Removed polling complexity
Music generation now uses
cassetteai-music
by default, which completes synchronously.
v2.1.0
Added async workflow for long-running music generation tasks.
Added
cassetteai-music
as a stable alternative for music generation.
v2.0.0
Migrated to ElevenLabs MCP server architecture.
Added voice cloning support.
v1.0.0
Initial release with TTS, music, and SFX routing.