Voice Cloning API
The most realistic voice cloning
Clone any voice from 5 seconds of audio. Instant cloning for rapid deployment, professional cloning for maximum fidelity. Every cloned voice runs on Inworld TTS with sub-200ms latency, emotion control, and 15-language support, ready for realtime applications at scale.
Original voice
5 sec sample
Cloned voice
Ready in minutessub-200ms latency
Clone any voice, three ways
Instant cloning for rapid deployment and user-generated voices. Professional cloning for maximum fidelity where it matters most.
Instant voice cloning
Create a custom voice from 5 to 15 seconds of audio. Zero-shot, available to all users in the Portal. Upload or record a sample and start generating speech in minutes.
Voice Cloning API
Automate voice creation programmatically. Let your users clone their own voices during onboarding, or batch-create voices for content workflows. JavaScript and Python examples on GitHub.
Professional voice cloning
Fine-tuned from 30+ minutes of clean audio. Recommended for uncommon voice types: children's voices, unique accents, or brand voices where instant cloning may not capture the full characteristics. Available by contacting the Inworld team.
Built for realtime applications
Every cloned voice runs on Inworld TTS — with the same production-grade latency, expressiveness, and multilingual support.
Production-grade latency
<200ms median first chunk for TTS-1.5 Max. <100ms for TTS-1.5 Mini. Streaming-native over WebSocket. No buffering delay.
Expression and emotion control
Audio markups for emotion ([happy], [sad], [angry]), delivery style ([whispering], [laughing]), and non-verbal vocalizations ([breathe], [cough], [sigh]). SSML break tags for precise silence placement.
15 languages
English, Spanish, French, Korean, Chinese, Hindi, Japanese, German, Italian, Dutch, Polish, Portuguese, Russian, Arabic, and Hebrew. Native-speaker quality in every language.
Half a cent per minute
$5/1M characters with TTS-1.5 Mini. $10/1M characters with TTS-1.5 Max. On-prem deployment for enterprises with data sovereignty requirements.
Full breakdown
| Feature | Instant Voice Cloning | Professional Voice Cloning |
|---|---|---|
| Audio required | 5 to 15 seconds | 30+ minutes (5 min minimum, 20+ recommended) |
| Availability | All users via Portal and API | Contact sales |
| Best for | Most applications, rapid prototyping, user-generated cloning | Uncommon voice types, brand voices, maximum fidelity |
| Supported formats | WAV, MP3, WebM (max 4MB) | Contact sales |
| Languages | 15 languages | 15 languages |
| Emotion and audio markups | Supported | Supported |
| Timestamp alignment | Word, character, phoneme, viseme | Word, character, phoneme, viseme |
| On-premise deployment | Available (H100/B200) | Available (H100/B200) |
| Zero data retention | Available | Available |
Use cases
Social & Tech
AI Companions and Social Apps
Persistent voice identity across sessions. Your companion sounds the same every time a user comes back.
Education
Language Learning
Clone instructor voices for consistent tutoring experiences across 15 languages.
Creative
Content Production
Produce podcasts, narration, and video voiceovers in your own voice without being in the booth.
Interactive
Gaming and Interactive Media
Clone character voices for dynamic in-game dialogue. Scale voice production without booking studio time for every line.
Impact
Accessibility
Preserve a person's voice for text-to-speech interfaces. Clone from a short sample and give users a voice that sounds like them.
How Inworld compares
| Inworld | ElevenLabs | Cartesia | |
|---|---|---|---|
| Min audio to clone | 5 seconds | ~60 seconds | 3 seconds |
| Latency (first audio) | <200ms (Max), <100ms (Mini) | ~300–400ms | ~40ms (base model) |
| Cost per 1M characters | $5 (Mini) / $10 (Max) | $11–$99 | $15 |
| Languages | 15 | 29 | 17 |
| Quality ranking | #1 Artificial Analysis TTS Arena | — | — |
| On-premise deployment | ✓ H100/B200 | ✗ | ✗ |
| Zero data retention | ✓ | Enterprise only | ✗ |
| Emotion & audio markups | ✓ | Partial | ✗ |
Latency measured as median time-to-first-audio. Cost as of March 2026. Quality ranking from Artificial Analysis TTS Arena independent blind evaluation.
FAQ
AI voice cloning creates a digital replica of a specific voice from audio samples. The cloned voice can then generate new speech that sounds like the original speaker. Inworld offers instant voice cloning from as little as 5 seconds of audio, and professional voice cloning from 30+ minutes for maximum fidelity.
Open the TTS Playground in the Inworld Portal. Click Create Voice, then Clone. Upload or record 5 to 15 seconds of audio. Your cloned voice is ready to use in the Playground or via API within minutes. For automation, use the Voice Cloning API with JavaScript or Python.
Instant cloning: 5 to 15 seconds. Professional cloning: 5 minutes minimum, 20 to 30+ minutes recommended. For instant cloning, record in a quiet environment and speak with varied emotion to capture the full range of the voice.
Yes. You must confirm you have the rights to clone the voice during the creation process.
All 15 supported languages: English, Spanish, French, Korean, Chinese, Hindi, Japanese, German, Italian, Dutch, Polish, Portuguese, Russian, Arabic, and Hebrew. Voices perform best when the synthesized text matches the language of the original sample.
Instant cloning (zero-shot) creates a usable voice from 5 to 15 seconds of audio. Available to all users. Professional cloning fine-tunes a model with 30+ minutes of audio for maximum fidelity. Recommended for uncommon voice types where instant cloning may not capture the full characteristics.
Yes. Automate voice creation programmatically. Useful for onboarding flows where users clone their own voice, or batch workflows. Code examples in JavaScript and Python on GitHub.
Cloning itself is free. You pay for speech synthesis: $10/1M characters with TTS-1.5 Max, $5/1M characters with TTS-1.5 Mini.
Yes. Audio markups for emotion ([happy], [sad], [angry]), delivery style ([whispering], [laughing]), and non-verbal vocalizations ([breathe], [cough], [sigh]). SSML break tags for precise silence. Emotion markups are currently experimental, English only.
Start building
Join millions of developers building the next wave of AI applications.