Method
The method
Heidi is built on what the research actually shows, rather than on what sells well as a language course. That leads to a few decisions that look strange at first.
- 01
The trap Heidi gets you out of
You learn German, move to Zurich, and find it does not help. Dialect is spoken at the table, you understand almost none of it, and because it shows, everyone politely switches to Hochdeutsch or English. The very input that would make you better is withdrawn because you need it. Heidi is a source of dialect that does not switch away.
- 02
Exposure beats rules
In the largest study of how people understand closely related languages, sheer amount of exposure mattered more than any measure of linguistic distance. Grammar is not what decides it; how much you have heard is. So Heidi is not a course of lessons but a place where real dialect keeps arriving.
- 03
Rules belong inside practice, not before it
Chind, Huus, isch, guet — the sound rules are real and useful. But the only clean test of teaching them as a lesson up front showed no measurable effect. What does demonstrably work: telling someone what to listen for, right before they hear it again. So Heidi shows one rule at a time, always next to a concrete word.
- 04
The test is always a new voice
Getting used to one speaker is easy and proves nothing. What counts is whether it carries over to a voice you have never heard. So Heidi trains with many speakers and always tests with an unfamiliar one.
- 05
Speaking comes last, and that is not a gap
Adults rarely reach native pronunciation in a second dialect, and in Switzerland that matters less than almost anywhere: understanding dialect and replying in Standard German is normal and respected. So Heidi does not sell you listening practice as a fix for your speaking — the evidence for that is weak.
Why this is not a detail
Bernese, Basel and Eastern Swiss forms are perfectly correct words — just not here. Someone learning Zurich German cannot, by definition, hear the difference. Which is exactly why that decision must not sit with a language model.
- tüütsch / tütschOstschweiz
- nidOstschweiz→ nöd
- güetBern→ guet
- gäuBern→ gäll
- …öu…Bern
- saiBasel
- ß→ ss
A fixed list of rules — not a language model. It checks every line Heidi shows you. You can run the list yourself here.
Zurich German has no official spelling. This check never tells you your spelling is wrong — only that a form comes from another region.
The loop
- 01You receive something you do not understand.
- 02Heidi explains it immediately — completely, not as a puzzle.
- 03A word or two sticks, because it was explained when you needed it.
- 04The same words come back later, in a different sentence.
- 05Eventually you meet them out in the world, and Heidi is not there.
The last one is the goal. Most programmes want you to come back. A learning product should want you to need it less.
What the research says
Language-learning products accumulate pseudoscience because “there is a study” turns very quickly into “this is proven” and then into a whole product. We keep three things apart: what is established, what we suspect, and what is simply a decision.
Established
We rely on these.
Exposure beats linguistic distance.
Across 1,833 listeners and 70 language pairs, exposure to the test language mattered more than lexical, phonological or orthographic distance.
Training with many voices is what carries over to unfamiliar ones.
Practising with a single voice can score better on that voice and fails to transfer. Confirmed specifically for regional dialects.
- Lively, Logan & Pisoni 1993. Training Japanese listeners to identify English /r/ and /l/. II: The role of phonetic environment and talker variability. Journal of the Acoustical Society of America 94(3), 1242–1255.
- Clopper & Pisoni 2004. Effects of talker variability on perceptual learning of dialects. Language and Speech 47(3), 207–239.
Saying what to listen for is an active ingredient, not decoration.
Same material, same feedback: only the group cued to the relevant contrast learned it.
Retrieval with feedback beats rereading.
222 studies, 48,478 learners; g ≈ 0.50, and 0.54 with feedback against 0.37 without.
Spaced practice beats massed, and the lead grows over time.
g ≈ 0.76 immediately, g ≈ 1.15 after a delay, across 48 experiments and 3,411 people.
Captions help — after the listening attempt, not during it.
Large effect on vocabulary (g ≈ 0.87), apparently because text helps cut the stream of sound into words. Permanently visible text becomes a crutch.
Listening training improves your own speaking only weakly.
d ≈ 0.92 for perception, d ≈ 0.54 for production, with no correlation between the two.
Writing dialect is digitally normal in Switzerland, not slang.
That is why “write like someone from here” is a real competence and not a gimmick.
Hypothesis
Plausible, untested — and Heidi is the instrument.
Consonant rules may predict intelligibility better than vowel rules.
What is established is that phonetic distance predicts intelligibility better than lexical distance. The precise figures this page once used to set consonants against vowels are in no source we could open — so the claim sits here rather than under Established. Two of our four front-page rules are vowel rules, and so the weaker bet either way.
Sound rules work as a cue inside practice even though they fail as a lesson.
The only clean test of the lesson form — 50 minutes of Dutch–Frisian — showed no significant effect, and the authors themselves warn against generalising it. The entire European intercomprehension tradition is, in the words of the leading researchers, essentially unevaluated. Our version is therefore the untested one. So we measure it.
A short tuning session measurably improves comprehension of an unfamiliar voice.
What is established after about a minute is faster processing — not more words understood. So we do not claim that a minute makes you understand more.
Decision
Product choices that stay right even if the hypothesis does not hold.
- Listening before writing before speaking — justified by the language situation, not only by evidence.
- Measured rather than gamified. No streaks, no points.
- The test voice is always one you have not heard.
- Real Zurich recordings, because every available Zurich corpus is licensed for research only.
- The model never judges its own dialect.
Where we corrected ourselves
This site once said the lesson form of the sound rules had been “tested and did not work”. A single 50-minute study does not carry that weight, and it made our own version look evidenced when it is the untested one. It also said there was no purchasable Swiss German speech synthesis; that is no longer true.