Clone any voice from a short audio sample and use it for text-to-speech generation.

Overview

Voice cloning creates a custom voice from a reference audio sample. Use it to:
  • Personalize TTS — Generate speech in a specific voice
  • Create characters — Unique voices for games or media
  • Accessibility — Preserve a person’s voice
  • Localization — Maintain voice consistency across languages

Getting Started

Download a Voice Cloning Model

Clone a Voice

  1. Prepare a reference audio file (5-30 seconds of clear speech)
  2. Use the voice for TTS generation

Using the Web UI

Voice cloning now lives inside the unified Voices workspace.

Step 1: Upload Reference Audio

  1. Navigate to Voices in the sidebar and choose the clone flow
  2. Upload a reference audio file
  3. The audio should be:
    • 5-30 seconds long
    • Clear speech, minimal background noise
    • Single speaker
The reference transcript is marked optional: Chatterbox clones from the audio alone, while Qwen3-TTS Base, VibeVoice, Fish Audio S2 Pro, Fish Audio Drama 3 and Maya 2 Native need it. With Maya 2 Native, the dialog also asks for the language of the recording, or detects it from the transcript; the voice can then speak any of the model’s languages.

Step 2: Generate Speech

  1. Enter the text you want to speak
  2. Click Generate
  3. Listen to the output in the cloned voice

Step 3: Save and Reuse

  • Download generated audio
  • Save the voice profile for future use

Using the CLI

Use izwi tts with a Base, VibeVoice, FishAudio-S2-Pro, FishAudio-Drama-3 or Maya2-Native-2.5B TTS model and provide both the reference audio and its transcript:
For longer reference transcripts, store the transcript in a file:
Chatterbox-500M and Chatterbox-Multilingual-500M clone from the reference audio alone. They do not condition on the reference’s words, so a transcript is optional and ignored. With the multilingual model, a reference in a language other than the request’s tends to lend the output its accent; Resemble AI suggests a matching reference, or cfg_weight 0 to soften the accent:
Maya2-Native-2.5B clones from reference audio and its exact transcript, and the cloned voice can speak any of the model’s 12 languages, not only the reference’s. Name the language of the reference with --reference-language (a code such as en), or leave it to be read from the transcript’s script. The reference must last 0.5 to 30 seconds; Maya Research recommends 3 to 10. Here an English reference speaks Hindi, which Izwi reads from the text’s Devanagari script (--language hi says the same):
FishAudio-Drama-3 clones from reference audio and its transcript, but the reference is optional: without one it speaks in a voice its sampling seed sets (see Text-to-Speech). A reference sent with the request may last up to 60 seconds; a saved voice holds up to 30 seconds, as for every model. For a dialogue it takes a reference per speaker, up to 8 in one request, each a clip with its transcript or a saved voice with one (--speaker-reference, --speaker-voice; references on the API); see Dialogue.
Saved voices created in Voice Studio or through /v1/voices can be reused from the CLI:
--saved-voice-id is mutually exclusive with direct --reference-audio and --reference-text input, and --reference-text needs --reference-audio.

Using the API

Endpoint

Request (JSON)

Example (curl)

Saved voices can also be managed through /v1/voices and reused from /v1/audio/speech. See the API Reference for the saved voice routes and exact fields. A voice saved with a reference_language passes it to Maya2-Native-2.5B requests that use it, unless the request sets its own.

Reference Audio Guidelines

Ideal Reference Audio

Good Examples

  • Podcast clips
  • Interview segments
  • Voice memos
  • Audiobook excerpts

Poor Examples

  • Music with vocals
  • Multiple speakers
  • Heavy background noise
  • Very short clips (under 3 seconds)
  • Whispered or distorted speech

Tips for Best Results

  1. Quality over quantity — A clear 10-second clip beats a noisy 30-second one
  2. Natural speech — Avoid monotone or exaggerated delivery
  3. Match content — Reference emotion should match desired output
  4. Consistent volume — Avoid clips with volume changes
  5. No music — Background music interferes with cloning

Available Models

Larger models produce more accurate voice clones.

Ethical Considerations

Voice cloning is a powerful technology. Please use it responsibly:
  • Get consent — Only clone voices with permission
  • Don’t impersonate — Never use cloned voices to deceive
  • Respect privacy — Don’t clone voices without authorization
  • Legal compliance — Follow applicable laws and regulations
Both Chatterbox models mark every utterance with Resemble AI’s imperceptible PerTh watermark, which Resemble AI’s open-source perth package can detect.

Limitations

  • Accent accuracy — May not perfectly capture all accents
  • Emotional range — Cloned voices may have limited expressiveness
  • Unique characteristics — Some voice qualities are hard to replicate
  • Language — Best results in the model’s primary language

See Also