Same gauntlet line read by Parler-mini and Irodori-MLX. Tap through top to bottom — the story is in your ears, the numbers are just the receipts. CER = character error rate vs the script (lower = cleaner).
Single pass, whole passage. Clean for ~15 seconds, then loses the thread — the back third dissolves into gibberish.
listen for: the ending falling apart
Parler-minirescued
Same model, fed one sentence at a time and stitched. Each chunk stays in the clean zone, so the breakdown disappears.
listen for: the ending now intact
Irodori-MLXclean, one pass
Different model, native on the M4. Handles the whole passage in a single pass — no chunking trick needed.
CER 5.3%27.7s · ~15s to generate
Irodori-MLXbest of 5 · the keeper
Five takes, different seeds, kept the cleanest. Same model and settings — just the best roll. This is the one worth banking.
CER 4.4%best of 4.4–8.8% spread
Irodori-MLXover-engineered
Per-sentence best-of-5, stitched. I expected this to win — it lost. Short fragments render choppier and the prosody seams show.
CER 8.8%worse than just picking a good whole take
The rule that fell out: chunking is a rescue for a model that breaks on long passages (Parler), not an upgrade for one that doesn’t (Irodori). When the model can already do the whole line, generate a few whole takes and keep the best — don’t chop it up.