Custom Voice Tester 🎛️

Each entry is a precomputed weighted recipe made from two or more existing Kokoro style embeddings. The entries are not 66 separately trained models, verified human identities, licensed celebrity voices, or objective quality winners. Names and trait labels describe the recipe's intended listening direction; they do not guarantee that every sentence will exhibit those traits.

The dominant source prefix selects a language route for phonemization. That is an implementation choice, not evidence that a blend has learned another language or accent. Cross-language and “fusion” recipes may be less stable than same-language sources. Count, metadata, and availability are tied to the matching experimental asset bundle rather than the public Kokoro base-voice catalog.

Compare blends with one fixed passage, model precision, speed, backend, browser, and device. Include names, numbers, punctuation, questions, and longer clauses, then replay the base voices that contribute the largest weights. Listen for pronunciation, rhythm, intelligibility, loudness, artifacts, and whether the supposed difference remains across more than one sentence.

Play All is a convenience audition, not a blinded benchmark: order, listener fatigue, loudness, and cached results can influence preference. Record the recipe and environment if a comparison will support a public claim. Do not label a blend “best,” assign a numeric quality score, or recommend commercial use without a disclosed corpus, multiple samples, and review of upstream voice and dataset terms.