Where we start
A bench is a fixed set of speakers and a fixed set of phrases, so that when two scoring methods disagree, the only thing that differed was the method. The speakers here are synthetic: each is a hum whose pitch follows the right tone shape, at a voice height and range inside the usual adult ranges, not a recording of a person. That makes the bench repeatable, and it makes one thing very clear. Every number below is computed by the same tracker that will listen to you.
★ What you walk out with tonight. A calibrated voice profile, your score on the twelve words of session 1 judged by shape and not by height, the counted reason a hertz-based checker fails lower voices, and one tone you now know you miss.
How a machine hears a tone
6 min liveThe tracker does five things, in order. It slices the sound into short windows, finds the pitch in each window, groups the voiced windows into syllables, turns each syllable into a shape on Chao's five-level scale, and picks the tone whose shape is nearest. Only the second step needs real mathematics, and it is a published method called YIN.
LiveWhat YIN does, in one paragraph3 min▶
A voiced sound repeats itself: the wave that leaves your throat on a steady pitch has the same shape every few milliseconds. YIN, published by de Cheveigné and Kawahara in 2002, slides a copy of each short window against itself and looks for the smallest delay at which the copy lines up with the original. That delay is one period of the voice; one divided by it is the pitch. A window with no clear repeat, such as a breath or the hiss of an s, has no pitch and is marked unvoiced, which is how the tracker finds where syllables begin and end.
Self-studyWhat the trainer cannot hear2 min▶
It scores pitch shape and nothing else. It cannot tell a q from a ch, an an from an ang, or whether a syllable was too long; it does not know which word you meant. A real voice also adds creak at the bottom of a third tone and small pitch bumps from consonants, which the synthetic bench voices do not have, so the bench claims only the ordering of the methods, never an accuracy for real speech. Use the trainer for tones, and a teacher or a native speaker for everything else.
"Your pitch is too low." A constructed case with a familiar shape. A Lumen engineer with a deep voice tried a free tone app that compared his pitch with a recorded female teacher's, in hertz. It marked nearly every fourth tone wrong because his "high" start was lower than her middle. He concluded he had a bad ear. His shapes were fine; the app was measuring height. The bench below counts what that method does to a lower voice.
Four voices, five ways of scoring, then yours
★ 26 min · the main eventThe table is computed when the page loads. Each synthetic voice says all twelve phrases with the tones a native speaker uses, and the table counts how many of the 23 scored tones each method hears as intended; then each voice says every phrase again with one tone deliberately wrong, and the table counts how many of those 12 mistakes each method catches. Read across a row to see what a method does to different voices; read down a column to see what changing the method does to one voice.
Read the lower voice's column first. The first row is the method many free apps use, and it is the row where a deep voice saying every tone correctly is told it got most of them wrong. Then calibrate and try your own voice below, with the twelve words from session 1.
Shape, not height
5 min liveThe bench above computes these live; the numbers here are the ones it produces, written down so they can be argued with after class. Change a speaker or a phrase in the script and they move. The ordering does not.
| Scoring method | Higher voice | Lower voice | Narrow range | Wide range | Mistakes caught |
|---|---|---|---|---|---|
| Raw pitch in hertz, one template voice | 23 | 6 | 6 | 23 | 12, 7, 7, 12 |
| Semitones from your own middle | 23 | 23 | 19 | 23 | 12 of 12 each |
| Range-scaled, calibrated (the trainer) | 23 | 23 | 23 | 23 | 12 of 12 each |
| Same, judging each phrase on its own | 21 | 21 | 21 | 21 | 12 of 12 each |
| Calibrated, but dictionary tones | 18 | 18 | 18 | 18 | 12 of 12 each |
Right tones heard right, of 23 scored; mistakes caught, of 12. Computed in your browser on four synthetic voices; with background noise added the trainer's method still scores 115 of 115 across five runs.
LiveFinding 1: hertz punishes the lower and the narrower voice2 min▶
A method that compares raw pitch with one template voice is right for voices near that template and wrong for everyone else. On the bench it scores the higher voice 23 of 23 and the lower voice 6 of 23, though both said exactly the same shapes, and it catches only 7 of the lower voice's 12 real mistakes. Measuring each syllable in semitones from the speaker's own middle fixes the lower voice. Scaling to the speaker's own range as well fixes the narrow-range voice, the person who speaks with little pitch movement, who otherwise loses 4 tones.
LiveFinding 2: a word on one level needs your profile2 min▶
公司 (gōngsī, company) is two first tones: high and level, twice. Judged on its own, a phrase with no pitch movement has no range to measure, so the method cannot tell "high and level" from "middle and level", and both syllables are missed. Calibrating once, on mā má mǎ mà (the four tones on one syllable), gives the trainer your middle and your range, and every phrase after that is judged against them. It is the difference between 21 and 23, and it is why the practice tool asks for calibration first.
Scoring the dictionary tones marks fluent speech wrong
4 min liveThe obvious way to build a checker is to look up each character's tone and score against it. The bench counts what that does: five syllables a native speaker says correctly, every first syllable of a tone change, are marked wrong, and the learner is pushed toward the careful, slightly foreign pronunciation of each character said alone.
LiveWhy the obvious checker is the wrong one2 min▶
A learner who is marked wrong for saying ní hǎo will try saying nǐ hǎo with two full third tones, which is harder to say and sounds stilted to a native listener. The checker's error becomes the learner's habit. The trainer applies the three tone-change rules from session 1 to the phrase first, then scores against the tones a native speaker actually produces; the five syllables the dictionary checker rejects are exactly the ones it accepts.
Self-studyThe honest limit: when dictionary tones are what you want2 min▶
Learning a new word, you need its dictionary tone: that is what you look up, and it is what you will see written in standard pinyin. The trainer's practice list shows both, the dictionary tones and the tones as said, for every phrase. The rule is simple: learn the word by its dictionary tone, say the phrase by its spoken tone, and score the phrase by the spoken tone.
The careful greeter. A constructed case. After a month with a dictionary-tone app, a Lumen manager greeted the Shanghai team with a slow, perfectly dipped nǐ, then hǎo. It was understood at once and sounded like a recording. Nobody corrected him; the Shanghai side switched to English, which is the quiet signal that a learner sounds effortful. One evening on the surface rules, and his 你好 (nǐ hǎo, hello) sounded like everyone else's.
Before session 4
30 min- Calibrate, then say the twelve words on three different days. Write the score each day and the one tone you miss most.
- Drill your weakest tone in pairs: if it is the third, alternate 买 (mǎi, buy) and 卖 (mài, sell); if the second, 十 (shí, ten) against 是 (shì, to be, yes).
- Read the five tone changes aloud until the changed tone comes without thinking: ní hǎo, láobǎn, bú shì, yídìng, yìqǐ.
- Bring a price and a date from your own work. Session 4 is numbers, money, dates and times, where tone mistakes become wrong quantities.
Three questions
Clear all three and session 3 gets stamped in your passport.
1 · A tone app compares your pitch in hertz with a recorded teacher's. Your voice is much lower than hers. What does the bench predict?
Tones are shapes relative to your own voice. A method that measures height against one template voice confuses "a lower voice" with "a lower tone", which is why "your pitch is too low" is never useful tone feedback.
2 · Why does the trainer ask you to say mā má mǎ mà before anything else?
Two first tones are level and high; with nothing else to compare against, "high" cannot be told from "middle". Your calibration is that comparison. Nothing is uploaded: the recording stays in the page.
3 · A checker scores every phrase against the dictionary tone of each character. What goes wrong?
Learn a word by its dictionary tone; score a phrase by its spoken tone. The trainer applies the three rules first and shows you both, so you see the change rather than being marked down for making it.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Session 3 cheat sheet · pin this
Tones carry the most weight in numbers. 十 (shí, ten) and 四 (sì, four) differ in tone and initial, and a price misheard by a factor of two or a date off by a week costs more than any greeting. Session 4 is numbers, money, dates and times, including 万 (wàn, ten thousand), the unit that makes Chinese prices read differently from English ones.