Armin Ronacher extracts a voice from 28 seconds of his own speech, mixes it with a generated voice, and adds emotional direction. His Qwen3-TTS experiment starts with a practical goal: giving a family dungeon crawler a distinctive narrator.

Voice identity becomes an editable input

Qwen’s speech models divide the work. Base clones a reference recording. VoiceDesign creates a voice from a description. CustomVoice supplies built-in speakers with instruction-controlled delivery.

Ronacher connects these paths through a speaker embedding: a numerical representation of vocal identity. Base produces 2,048 values for each reference. He blends two vectors, rescales their length, then generates speech with the result. Saving that speech creates a reusable reference, although extracting its embedding again changes the voice slightly.

The more surprising step is passing Base’s voice representation into CustomVoice’s lower-level generation function. That lets him pair a cloned or blended identity with instructions for sadness, anger, excitement or whispering.

Qwen’s official documentation separates Base cloning from CustomVoice instruction control. Ronacher’s bridge uses private Python helpers and qwen-tts 0.1.1. It should be treated as version-sensitive experimental code.

A synthesis representation has limits

Ronacher reports weaker emotional control for cloned voices than for built-in speakers. Recording conditions also affect the result: a conference microphone leaves different characteristics from a clean synthetic voice.

His speaker-identification side experiment exposes another limit. A conference reference fails to identify him reliably in a podcast using Qwen’s embeddings. A reference from that podcast gives about 98% agreement. ECAPA, a dedicated speaker-verification model, reaches about 97% using the conference reference. Those are his small-sample results, not a general accuracy ranking.

The article also shifts perceived voice characteristics along a direction derived from generated male and female references. That is an experiment on a particular sample set, not a universal model of gender.

Caveat: We have not reproduced the synthesis or evaluated the audio; the original provides interactive samples and code. Use recordings with permission, and do not assume that blending makes a person’s voice anonymous.