Speaker identity is split across conditioning sites

Audio samples for the paper by Callum Macpherson (2026). Code and data: github.com/CallumJMac/conditioning-sites.

Each item below has the prompt the systems were given, a different recording of the same speaker (the held-out reference used for SIM-o), and each system's output with the prompt ("full") and with the prompt removed but the speaker embedding kept ("x-vector only"). The last table for each item is the CosyVoice-300M factorial, which switches the prompt and the two embedding sites on and off independently.

Items were drawn at random with a fixed seed (scripts/build_samples.py), not picked to illustrate the result. All voices are synthetic imitations of LibriTTS-R and Common Voice speakers. XTTS-v2 and F5-TTS outputs are licensed for non-commercial use only.