Armin Ronacher digs into Qwen3-TTS: clone a voice from 28 seconds, then blend embeddings

Armin Ronacher · rss · 2026-10-29

Armin Ronacher shares a hands-on deep dive into Qwen3-TTS, used to narrate a dungeon-crawler game he's building with his kids:

Model split: three 1.7B variants — Base (clones from seconds of reference audio), CustomVoice (nine built-in speakers + instruction control), VoiceDesign (invents a voice from a text description). Officially, voice cloning and instruction control don't combine.
Low cloning bar: 28 seconds from his PyAI Conf talk (transcribed with Whisper) suffices for a convincing clone; clean studio recordings aren't required.
Embedding tricks: Base's speaker embedding is a 2048-dim vector you can do arithmetic on — a weighted average of two voices yields a new in-between voice, which can be baked back into a reference clip.
Breaking the instruction limit: since Base and CustomVoice share nearly identical transformer weights, you can feed Base's x-vector into CustomVoice's generate call to get cloned voices that follow emotional instructions ("sad, on the verge of tears").
Speaker ID caveat: embeddings as a speaker detector fail with the conference clip as reference (98% accuracy using a 20-second podcast intro instead), while ECAPA hits 97% even on the conference clip.

Code uses the official qwen-tts package (MPS supported); examples are reproducible.

More from Multimodal