Code Switcher

A text-to-speech system that infers who is speaking from how the sentence code-switches, rather than cloning a voice you hand it. Trilingual ES–EN–ZH. Wizard-of-Oz build: Stages 0–1 are real, Stage 2 is hand-authored.

z_spk ~ p(z | Φ(text)) audio ~ p(a | text, z_spk) — no reference-audio port anywhere

Input utterance

Guidance scale1.00

Low: an idiosyncratic individual who happens to write this way. High: the archetype. The entropy floor is the caricature governor.

Seed0

Same sentence, different plausible person. The posterior has width, so identity is sampled, not looked up.

What is real here and what is not

Real: tokenization, two-pass language ID, frame-vs-mass dominance, switch typology, orthographic and discourse-marker extraction, register.
Hand-authored: the feature → speaker map in Stage 6. Weights are estimates, not learned. That is the point of a Wizard-of-Oz: test the inference before spending a training run.
Not speech: Stage 7 renders a parametric signature of the speaker vector so two inferences can be compared by ear. A real backend needs a TTS with a speaker-embedding port that accepts an off-reference vector.
Known gaps: Han segmentation is character-level; romanized Mandarin (aiya, hen hao) is a stub list; no POS tagger, so switch typology approximates constituent structure from closed-class membership.