How Familect works, and where it's honest about its limits

Where the linguistics comes from

Every phonological field in Familect's parameter schema is backed by a citation to standard descriptive/reference literature, not invented by the engineering team. Familect's parameter model is a reconstruction and operationalization of this literature into something machine-checkable -- it is not itself a citable linguistic source.

• Lipski, John M. Latin American Spanish. Longman, 1994.
• Hualde, José Ignacio. The Sounds of Spanish. Cambridge University Press, 2005.
• Canfield, D. Lincoln. Spanish Pronunciation in the Americas. University of Chicago Press, 1981.
• Mateus, Maria Helena & Ernesto d'Andrade. The Phonology of Portuguese. Oxford University Press, 2000.
• Valdman, Albert. Descriptive work on Haitian Creole phonology and lexicon, including the Haitian Creole-English Bilingual Dictionary (Creole Institute, Indiana University).
• Kouwenberg, Silvia & Eric Murray. Papiamentu. Lincom Europa, 1994.
• Schwegler, Armin. Published work on Palenquero (San Basilio de Palenque) phonology and Afro-Hispanic creole studies.

125 attested varieties across 8 languages (English, Spanish, French, Haitian Creole, Mixed & Border, Palenquero, Papiamento, Portuguese), each traceable back to this reading list.

What the parameter matrix knowingly cannot model

The matrix is a flat set of parameters, not a full phonological grammar. Five gaps are documented explicitly rather than papered over:

Papiamento's lexical tone

Papiamento is documented as having genuine lexical/word-level pitch-tone contrasts. There is no tone-bearing-unit representation anywhere in the schema -- Papiamento presets cannot express lexical tone at all, in the scoring matrix or in the synthesized audio.

Córdoba (Argentina) pretonic vowel lengthening

The distinctive “tonada cordobesa” stretches the syllable before the stressed one. No dedicated field models pretonic lengthening specifically; the Córdoba preset approximates it with the engine's closest available proxy (phrase-final lengthening), not a faithful rule for this specific environment.

Caipira (rural São Paulo) retroflex R

The retroflex “R caipira” [ɻ] has no dedicated matrix value. The Caipira preset approximates it with the closest available IPA-adjacent option, which is articulatorily distinct from the true retroflex.

Yucatec Spanish final /n/

Yucatán Spanish's Maya-substrate-influenced final /n/ behavior is distinct from the Caribbean/Andalusian velarization the schema does model, and has no dedicated field of its own.

Porto (northern Portugal) betacism

The historic b/v merger is a real, marked feature of northern Portuguese speech. It is a lexical/segmental trait the schema has no dedicated field for, so it is documented in preset copy but not phonetically rendered.

The audio is synthesized, not recorded

Every clip you hear in Familect is synthesized, never a recording of a native speaker -- and where a play button offers a choice, there are two genuinely different engines behind it, not one engine with a quality slider. Phonetic (espeak-ng) is handed the exact IPA for the accent you're hearing and speaks it precisely, sound by sound -- the only way to accurately render an accent that was composed by hand and was never actually recorded by anyone, at the cost of a mechanical, synthetic voice quality that this app labels honestly rather than pretending otherwise. Natural, where available, instead hands a plain-letter respelling of the accent (Puelto Rico, not /pweɾ.to ˈri.ko/) to a human-sounding AI voice model -- it sounds like a real person, but it approximates the accent through spelling rather than exact phonetic control, so very fine sound distinctions land more reliably on phonetic. Neither is a strictly better version of the other; they trade phonetic precision for listening comfort in opposite directions, and this app shows you which one you're hearing rather than blurring the difference. Beyond the voice-engine choice itself, the accented IPA a preset produces passes through a second mapping layer before it becomes sound, and that mapping is either confirmed exact against a live probe, or a principled nearest-neighbor approximation.

Read live from the running synthesis engine: of 103 distinct IPA symbols the transformer can emit, 54 render exactly and 49 are approximated, with none unmapped.

One limitation no mapping choice can fix: the synthesis engine sets a single flat baseline pitch for an entire utterance -- it cannot produce a real melodic intonation contour. Treat any generated audio's prosody as decorative, not a faithful rendering of how a real speaker's pitch actually moves.

Grapheme-to-phoneme (g2p) engine coverage, by language

Text is converted to IPA by a different engine per language. This table is read directly from the dispatch registry the API actually uses, so it can't drift out of sync with what happens when you submit an attempt.

The score's validation status

Read this plainly: zero learners have completed a single exercise of any kind, and zero native speakers or phoneticians have rated Familect's scoring against real human judgement. The overall/accuracy scores you see measure how closely a recording matches one preset's specific, hand-set parameter choices -- a construct-validity claim, not a correctness claim, and one that has not yet been checked against how real listeners actually perceive these accents.

This matters most for the varieties this tool models that are stigmatized within their own speech communities. If you are a native speaker of one of these accents practicing your own family's variety, and the tool marks a feature of your natural speech as a “miss,” that is this tool disagreeing with one specific, curated parameter preset -- not a judgement that your speech is wrong. A validation study (recruiting native-speaker raters against real recordings) is planned but has not been run.