voice-cloning
Installation
SKILL.md
Voice cloning
Produce speech in a custom voice, one you either clone from a real recording or design from a text description, then synthesize any text in it. The difference from plain text-to-speech: the voice is bespoke, not picked off a preset list. Two intents live here, clone-from-sample and design-from-description, and they route to different models.
Inputs to collect
- Which intent: clone an existing voice (have a recording) or design a new one (have a description). This picks the model.
- For cloning: a clean reference recording of the target voice. A few seconds is enough (Qwen ~3s, Dia ~5-10s). Mono, low noise, single speaker. Plus a transcript of what's said in it if using the Qwen clone path.
- For designing: a voice description covering gender, age, timbre, accent, pace, and mood (e.g. "warm older male, gentle British accent, unhurried").
- The text to speak, and the target language.
- Consent / likeness: clone only voices you own or have explicit rights to. Refuse cloning a third party's voice without permission.
Models
- Clone from a sample (single speaker): Qwen3-TTS Base (
alibaba:qwen@3-tts-1.7b-base). Takes the reference recording ininputs.audiowithspeech.voice: "clone". Highest-similarity mode (ICL) needs atranscriptof the sample; embedding-only mode (settings.xVectorOnly: true) skips the transcript at lower similarity. 10+ languages. - Clone for dialogue / multi-speaker: Dia 1.6B (
runware:dia@1.6b) or Dia2 2B (runware:dia2@2b). Reference voices go ininputs.audios(first =[S1], second =[S2]), supports non-verbal cues. English only. Dia2 streams and is the newer pick. - Design from a description: Qwen3-TTS VoiceDesign (
alibaba:qwen@3-tts-1.7b-voicedesign). The voice is created from a natural-language description inpositivePrompt, withspeech.voice: "design". No recording needed. - Premium preset timbres (not cloning): Qwen3-TTS CustomVoice (
alibaba:qwen@3-tts-1.7b-customvoice). Nine curated timbres (aiden,serena,vivian, …) plus a style hint. Use it when a polished stock voice with style control beats a clone.