Romanian · Study sheets v1 · ZH / RO / EN
Hear the Difference: A Romanian Perception Pilot — Six Contrasts
This version is sheets only, no test: the letters, IPA and articulation of six contrasts are all here; discrimination is learnable only through listening tests, and the test waits on validated multi-talker human recordings — no synthetic voices, no single talker.
Sheets carry no audio. Sheets cannot teach discrimination (the page's own sentence, below). The design evidence chain is on the sources page, S1–S4.
Three sentences first: sheets teach letters, not discrimination; the six are editorial starter hypotheses, not your diagnosis; the site reports no learning.
Six contrast study sheets (no audio)
Sheets cannot teach discrimination; discrimination is learnable only through listening tests — the page's own sentence. Each sheet gives letters, IPA and articulation sketches only; no Chinese analogies, no production coaching.
Contrast 1 · a – ă (/a/–/ə/)
- Letters: a and ă (a with breve).
- IPA: /a/ open front-central vowel; /ə/ mid central vowel, lips relaxed.
- Articulation sketch: /a/ jaw open, tongue low; /ə/ mouth half-open, tongue mid-central, everything relaxed.
- Acoustics: the doua-note vowel table (fir/tren/cânt/casă/pas/nord/nor).
Contrast 2 · ă – î/â (/ə/–/ɨ/)
- Letters: ă and î/â (î and â are one sound /ɨ/; standard orthography splits them by position — check each word on dexonline).
- IPA: /ə/ mid central; /ɨ/ high central unrounded (tongue high and central, lips unrounded).
- Articulation sketch: /ɨ/ tongue body raised toward the mid palate, tip down, lips spread.
- Acoustics: the doua-note vowel table. Note: calling /ə/ “mid central” is a broad phonemic convention (fine debate exists, see sources).
Contrast 3 · i – î (/i/–/ɨ/)
- Letters: i and î/â.
- IPA: /i/ high front unrounded; /ɨ/ high central unrounded — same height, different front-back.
- Articulation sketch: /i/ tongue near the front palate; /ɨ/ same height, body pulled back to center.
- Acoustics: the doua-note vowel table.
Contrast 4 · s – ș (/s/–/ʃ/)
- Letters: s and ș (s with comma).
- IPA: /s/ alveolar fricative; /ʃ/ postalveolar fricative.
- Articulation sketch: /s/ tongue tip near the ridge, hiss forward; /ʃ/ tongue pulled back toward the front palate, hiss back, lips lightly gathered.
- Gap statement: no s–ș acoustics page exists on this site (doua-note does vowels only); no acoustics invented here.
Contrast 5 · t – ț (/t/–/t͡s/)
- Letters: t and ț (t with comma).
- IPA: /t/ stop; /t͡s/ affricate (one breath: block then hiss).
- Articulation sketch: /t/ tip on the ridge, burst and done; /t͡s/ same burst immediately followed by hiss, no break.
- Gap statement: no t–ț acoustics page exists on this site; no acoustics invented here.
Contrast 6 · r – l (/r/–/l/)
- Letters: r and l.
- IPA: Romanian /r/ (commonly an alveolar tap or short trill — trill NOT required); /l/ lateral.
- Articulation sketch: /r/ tip taps the ridge once or a few times; /l/ tip on the ridge, air around the sides.
- Gap statement: no r–l acoustics page exists on this site; no acoustics invented here.
Spelling channel: dexonline (each word's spelling and stress from there; this page publishes no word lists).
The paradigm: why the test looks like this (design, not promise)
- The canonical three (canonical HVPT): multi-talker + varied phonetic contexts + immediate corrective feedback. Numbers from the originals only: multi-talker immediate gains g=.28 (k=13), generalization to untrained talkers g=.36 (k=9, with heterogeneity + publication-bias sensitivity); canonical pre-post g=0.92 (k=96), treated-control g=0.67 (k=32), retention confirmed, generalization “to some extent”. Identification is common practice, not mandate; minimal pairs are OUR design choice, not an HVPT requirement; natural tokens are the founding program's practice.
- What is measured (construct statement): word identification under minimal-pair competition (listening + orthographic mapping), not pure phoneme perception.
- Scoring skeleton: 16 first-plays per contrast for ANY contrast-level statement (4 talkers × 2 pairs × 2 reps, randomized); 2AFC with the 50% chance line on every display; overall score = correct first-plays ÷ total first-plays (pooled, unweighted, formula as stated); replays counted separately, no penalty (penalty weights would be arbitrary); 0.6× slow play as post-response study aid only, trial voided for diagnosis (temporal-distortion warning, stops/rhotics worst); timeout = missing (excluded, counted; 2 timeouts → rest prompt); failure → repeat the block (max 2), then continue-with-flag (no hard locks); ≥80% (latest block) yields only a “ready to add a contrast” suggestion, never a mastery claim; reassessment via spaced repeats, no decay modeling; below 16 trials only 题目太少，暂时不能判断, no percentages.
- What the six are: editorial starter hypotheses (central vowels prominent in literature + sibilant/affricate/rhotic coverage), Mandarin-relevance = hypothesis/pilot until cohort evidence. Curriculum graduation rule: aggregate anonymous data (privacy-treated + consent) → revise/select; no data = hypothesis forever (dependency stated).
- V2 path (specified, not built): 5th talker + untrained words + pre/post + no-feedback evaluation (held-out).
The game would report session performance only (“on these trials you got X”), never learning; generalization discussed as prior research only. Japanese-/r–l/ evidence does not validate the Romanian-for-Mandarin pairing.
Recording status: pending validation, hence no test
- Validation gates (unpassed): 4 standard-RO talkers (2F+2M, consent + pay + accent history + age + reuse rights archived); fixed mic/distance/room/chain (archived); calibration tone; isolated words + one fixed carrier variant per pair (finalized with the phonetician); conversational rate; no teaching voice (exclusion rule); peak ≤−3dBFS, noise ≤−60dBFS, SNR thresholds; 150ms silences; level-only processing (no spectral changes); loudness matched within pair+talker (±1 LUFS) + peak checks; blinded validation by 2+ natives; rejection criteria (either-validator miss, acoustic outlier, cue artifact); native-ceiling pilot (≥90% on the final set or item revised); anti-cue verification (no systematic level/duration/silence differences); automated manifest agreement checks (word/key/audio/talker/IPA/ID); sha256 per file + manifest version.
- Why no synthetic voices: the protocol demands natural multi-talker material + archived consents; synthetics are single-voice synthetic, failing mandatory multi-talker + the natural-token requirement — that would be a different product, not shipped.
- Freeze rule: any contrast with <4 valid talkers or a failed ceiling → withdrawn. The page is now exactly in total-freeze: sheets work, the test does not open.
Method drawer: pairs, words, talkers, scoring, limits, version
- Pairs (rules first, list unpublished): identical syllable count; identical stress pattern with the target syllable stressed in BOTH members; identical target position (initial/medial/final); flanking segments class-matched (same manner minimum, same place preferred); no morphological alternation across members; no proper names; no archaic/regional/dialectal items; no unstable recent loans; no homographs; no function words with predictable reduction; frequency band matched (same quartile in the build-named RO corpus); imageability comparable (rated, not assumed). Near-minimal pairs only labeled “near-minimal (differs in X)” + same controls otherwise + diagnosed separately.
- Words: bank ≥3 pairs/contrast; block samples 2. Per-pair audit tables + transcriptions + phonetician sign-off — three unpassed gates, list unpublished.
- Talkers: 4 standard-RO (2F+2M) + documentation, see above.
- Scoring: contrast statements take first-play accuracy + Wilson 95% CI (n≥16) + session observations (with counts). Raw counts always visible. R1 multi-talker-mandatory; R2 observations-not-merges, no L1 assertions; R3 no-efficacy-promises; R4 waveform-as-aid-only, no acoustic assertions; R5 performance-not-learning; R6 statement thresholds (n≥16 + all-talkers + bank≥3).
- Limits: out of scope: production coaching (articulator about-lines only), vocab/grammar, spectrograms/formants, new contrasts without protocol, efficacy/transfer assertions, Chinese-phonology assertions, leaderboards/accounts, held-out evaluation assertions.
- Version: sheets v1 (2026-09-17). Any audio/pair change = new manifest version + dated note.
Questions
Will the sheets teach me to hear the difference?
No. Sheets teach letters (letters, IPA, articulation); discrimination only through listening tests; the test awaits recording validation. Sheets cannot teach discrimination — the page's own sentence.
Are the six my difficulties?
Unknown. The six are editorial starter hypotheses, not your diagnosis; Mandarin-relevance is hypothesis until cohort evidence. Aggregate anonymous data will revise/select the list; without data, hypothesis forever.
Why not ship AI voices as a stopgap?
They fail the protocol (natural material, multi-talker, archived consents, blinded validation, ceiling pilot). A “score” on synthetics would not diagnose human-voice discrimination — misleading, not shipped.
How do auzi, doua-note, dexonline relate?
auzi measures thresholds (we train phonemically, different mechanism; auzi itself is in preparation); doua-note explains vowel acoustics (we link only the three vowel contrasts, consonants/rhotic unlinked with the gap stated); dexonline is the spelling-and-meaning channel.