The four pitch accent patterns
Every Japanese word carries one of four pitch accent patterns, written in dictionaries as a number: the mora after which the pitch drops.
- Heiban, type 0, flat: low, then high all the way, and a following particle stays high. Example: ame ga, candy. No drop.
- Atamadaka, type 1, head-high: high on the first mora, then an immediate drop, and low after that. Example: ame ga, rain.
- Nakadaka, type 2 and up, middle-high: low, rise, then a drop in the middle of the word. Example: kokoro ga, heart, dropping after mora 2.
- Odaka, type equal to the word's mora count, tail-high: low, high through the word, and the drop lands on a following particle. Example: otoko ga, man, dropping on ga.
The type number is the dictionary convention. Dictionaries and apps like OJAD mark it as a small number next to the word: hashi [1] is chopsticks, hashi [2] is bridge, hashi [0] is edge. Only atamadaka starts high. If your first mora is high, the pattern can only be atamadaka; if it is low, atamadaka is already ruled out.
Why every word ends in ga
Odaka and heiban sound identical on the word alone: both are low, then high to the end. The difference appears only when something follows the word. After a heiban word, a particle stays high. After an odaka word, a particle drops low. So every item in this app is practised as "word + ga": hashi ga, ame ga, otoko ga. The particle is the last mora the app checks, and it is what separates odaka from heiban. This is also how the pattern works in real speech: the pitch of what follows a word depends on the word's pattern.
Calibrate your voice range first
Before practising, press Set my voice range and slide slowly from your lowest to your highest comfortable note, saying a steady "ahhh" while you slide. The app measures your range from that slide and places the high/low line in the middle of it: a mora clearly above the line reads as high, clearly below as low. A mora landing near the middle is honestly reported as unclear rather than guessed. You can recalibrate any time in Settings, and your range is stored only on your device.
The 18 two-mora words
Six words per pattern family, built around the classic minimal sets: the same sounds, different patterns, different words.
- Heiban: ame candy, hashi edge, kumo spider, sake rice wine, kami paper, hana nose.
- Atamadaka: ame rain, hashi chopsticks, sake salmon, kami god, hana flower, neko cat.
- Odaka: hashi bridge, kumo cloud, kami hair, inu dog, natsu summer, fuyu winter.
Two sets cover all three patterns on one sound: hashi is chopsticks (atamadaka), bridge (odaka), and edge (heiban). Kami is god (atamadaka), paper (heiban), and hair (odaka). Practise each set in turn and the patterns separate fast.
Longer words: 12 three-mora words on the hard contrasts
Knowing two-mora patterns is only half the skill. Longer words add the nakadaka drop in the middle, and the heiban-vs-odaka contrast gets harder to hear as words get longer.
- Heiban: sakana fish, sakura cherry blossom, densha train, kodomo child. Ga stays high.
- Nakadaka: kokoro heart, sukoshi a little, taberu to eat, nageru to throw. Drop after mora 2.
- Odaka: otoko man, katana blade, namae name, agaru to go up. Drop on ga.
Longer words unlock at 50 points, after you have the two-mora patterns down. Say each mora on an even beat, like a metronome: the app splits your utterance into equal mora slices, so uneven morae confuse the check.
How the live check works
Your browser samples the microphone and measures your pitch with autocorrelation, a standard signal-processing method that finds how often your voice waveform repeats. No model is downloaded and no audio is sent anywhere. Your utterance is split into morae by equal time division, because Japanese morae are near-equal in length, and the median pitch of each mora is read as high or low against your calibrated range. The drop position is read straight off the high/low sequence. A mora landing near the middle of your range makes the reading unclear rather than guessed. Press Hear the target to hear any pattern as a stepped tone, one even step per mora, placed inside your own range before you say it.
Odaka vs heiban, and atamadaka vs everything else
Two confusions trip learners up most. Odaka vs heiban: identical without a particle, separated only by ga. Say the word alone and no one, including this app, can tell them apart. Atamadaka vs everything else: only atamadaka starts high. If your first mora is high, the pattern can only be atamadaka; if it is low, atamadaka is already ruled out. When one mora lands in the middle of your range, the app says so honestly instead of guessing, and names the difference to aim for.
Tokyo patterns, and what the app does not grade
The patterns taught here are standard Tokyo citation forms, the ones dictionaries list. Regional dialects differ: in Kansai, hashi chopsticks and bridge swap patterns. Real sentences reshape pitch further: downstep lowers later phrases, phrasing groups words together, and question intonation rises at the end. This app grades one word plus ga in isolation. It does not grade sentences, and it does not check grammar, vocabulary, or meaning. Devoiced morae carry little voicing and may not register: if a mora will not register, say it a little more fully voiced for practice.
Quiz, points, levels and badges
Quiz mode asks for 8 items in a random order: +10 points per match, +20 bonus for a perfect round. Levels run New Voice, Pitch Learner, Pitch Star, and Master of Japanese Pitch. Badges cover the first matched pattern, practising all 18 two-mora words, practising all 12 longer words, a perfect quiz round, and practising on seven days in a row. Confetti and chimes celebrate wins, and both can be turned off in Settings. Practice stays the core: the games are there to pull you back for one more pattern.
Your voice never leaves your device
Voice frames are analysed on your device and discarded. The page security policy blocks connections to every outside server. There is no account, no analytics, no recording, and no cookie. The built-in Offline proof panel shows a live log of every network request the page makes, lets you simulate a fully offline session, shows exactly what is stored on your device, and deletes it all with one button. The app works offline after your first visit. Pitch tracking approximates patterns, and regional variation exists: the target patterns describe standard Tokyo citation forms. Native-speaker review is being sought in parallel, not as a gate.
Learning tones too? Try the Mandarin Tone Trainer, the Cantonese Tone Trainer, the Vietnamese Tone Trainer, and the Thai Tone Trainer, the Taiwanese Hokkien Tone Trainer, and the Yoruba Tone Trainer, the same live pitch feedback for six tone languages. For sign language, Fingerspell teaches the ASL alphabet.
This is a home practice tool, not a teacher, therapist or safety device. Everything runs on your device: no video, audio or images are sent anywhere. Feedback is automatic and can be wrong. For real instruction, learn from a qualified Japanese teacher or a native-speaker community, linked on the Resources page.