How should WordSense pronounce Indonesian words out loud?

The alias strategy tested well on meja / boleh / gue. This batch pushes it further: 10 standalone words (mostly stored IPA already in the catalog, respelled with the same alias convention) and 5 short phrases that put an alias word inside connected speech, to see if the fix survives once the word isn't spoken alone.

Two words are deliberate controls: beli and senang. Both have a schwa vowel, which is the model's default reading anyway — so their "alias" is spelled identically to the plain word (no diacritic). They should sound exactly like the unmodified baseline. If a control sounds different, that's a sign the alias mechanism itself is doing something unexpected, not that a respelling rule is wrong.

Respelling convention used: é = /e/ in an open, non-final syllable (meja→médja). è = /e/ in a closed, word-final syllable (boleh→bolèh). A bare word-final open /e/ gets an added "h" digraph (gue→gu-eh). Plain "e" with no mark = schwa, left unmarked — that's the beli/senang controls.

WordSense already stores a per-word respelling that learners see on-screen right now (e.g. beli → “buh-LEE”) — the alias-batch above instead used a brand-new é/è convention invented for this test. If the exact string the app already displays also drives the voice correctly, display and audio share one source of truth forever, with no new convention to invent or maintain.

Convention v2 (updated from tonight's ear findings): é = full /e/ in any position, including word-final (témpé, not témpeh — you caught the -eh digraph reading as schwa on tempe). è = full /e/ in a closed final syllable (tèh, bolèh). Plain "e" = schwa, left unmarked. New this batch: a c→ch guard before e/i (chépat class) generalized to a second word (kecil→kechil), and no separators or stress marks in any alias string.

Unplanned finding — t070's stored IPA has a systematic bias, not just on cepat. Every candidate was cross-checked against Kaikki/Wiktionary's IPA (a second, independent source already in the catalog). Beyond cepat, seven more words disagreed the same way — t070 says a full, stressed vowel early in the word; Kaikki's Standard-Indonesian entry says schwa with stress later (dekat, penting, karena, lebih, rezeki, terus, berhenti). Those seven are brand-new words never ear-tested before, so they're respelled here from the Kaikki-corrected reading (identity alias, badged DATA-FIX) and t070's row is flagged as a data-fix candidate. Three words from the 17-Jul batch (kena, empat, sedih) show the same Kaikki disagreement — but you already ear-approved their é-marked version as "sounds right," so those were not silently changed; they're kept as-is and just badged ⚠ FYI for your awareness. dekat specifically matches a Kaikki-tagged Eastern-dialect variant rather than a flat error — a register question, not just a bug.