The one question: does the shipped engine bundle (tts-engine.js + shelf ONNX Runtime + Kokoro q8 graph), running in THIS real browser, produce clean (non-NaN) waveforms on long inputs? In the Node proxy, a ~450-token run returned an all-NaN waveform, and even after the shipped 140-token split one piece of the stress paragraph still NaN’d (107,400 of 699,000 samples). This page runs the shipped bundle as built — no stub, no reimplementation — through its own createEngine({modelDir, runtimeDir}).speak(text, voice, speed) entry, the same call the app makes.
The 86 MB model exceeds the per-file asset limit, so it is served as four chunks which this page concatenates before the engine asks for it. The engine fetches the model by URL exactly as in production; this page answers that one fetch from the assembled buffer (the same seam the app’s service worker occupies in production). Concatenation is the only transformation — the SHA-256 below must equal the registry pin or the battery will not run.
Assembling…
Voice af_heart, speed 1. Expected phoneme counts were measured with the branch’s own G2P in Node (gold+silver lists); if a run’s measured count differs from its expected count, the silver lists likely did not load in this browser (the engine then runs gold-only) and the run is still valid but not the same input — compare counts first.
| Run | Phonemes (expected) | Tokens | Samples | NaN samples | Peak | Time | Verdict |
|---|