Token-level MB-iSTFT-VITS · audio streams chunk-by-chunk (24 frames ≈ 384 ms) — first sound in ~100 ms. zh-TW + English. Runs entirely client-side (WASM).
No input needed — generates a fresh short tale and speaks it as it writes.
Ask anything — it replies in zh-TW as it generates, spoken clause-by-clause.