听力实验室 TingliLab.orgFree resource · help

help for the free tone lab · app v0.33.1 · October 2026 · back to the tone lab · privacy

The free tone lab is the part of TingliLab that costs nothing to run, so it costs nothing to use: no account and no limits. It drills the four tones by ear and by voice, with a tone judge that listens to your pitch; it gives you dictation on the HSK 1–3 words, in pinyin or 注音 (zhuyin), with Mainland or Taiwan readings; and it teaches you to write the characters, stroke by stroke. Everything you do stays in your browser, on your device; recordings are never uploaded. It runs in the browser, installs to a phone like an app, and works offline once you've prepared a voice. It is not a course: use it alongside whatever you use for vocabulary and reading.

Getting started

Open tinglilab.org in Chrome or Safari, on a phone or a computer. There is nothing to sign up for. The first time, a short welcome says what's here; Start with tone dictation starts a round straight away.
To keep it on a phone, install it: in Safari, Share → Add to Home Screen; in Chrome, the menu → Install app (or Add to Home screen). It then opens like an app. Settings → This site says how on your device.
To practise offline, prepare a voice first: Settings → Voices → Prepare this voice (see Offline).
Your practice log stays in this browser only. Download a backup now and then (Settings → Data): clearing the site's data in your browser removes everything, and there is no copy anywhere else.
If you read zhuyin, or learn Taiwanese Mandarin, set that up first: see Readings and standard.
The header's ? opens this page, and the version comes last. If something looks odd after an update, reload the page once.

Readings and standard

Two settings in Settings → Readings shape the whole tone lab.

Readings: pinyin, 注音 zhuyin, or pinyin + 注音. Whatever you choose is used everywhere a reading appears: in the tone drills' reveal and four-tones row, in the dictation reveal and in Write. Zhuyin follows Taiwanese conventions: first tone unmarked, the neutral-tone dot ˙ before the syllable. Tone colours colours readings by tone.
Standard: Putonghua (Mainland) or Guoyu (Taiwan) · provisional. Guoyu rotates the Taiwan voice in Tones, Dictation and Write (Settings → Voices), and, for what a Taiwan voice says, accepts the full tone wherever Putonghua uses a neutral one (谢谢 as 4·4 as well as 4·轻) in the tone quiz and in dictation. The Taiwan readings of the HSK 1–3 words were checked against Taiwan's Ministry of Education dictionary.
Each word is marked by the standard of the voice that says it. When a Taiwan voice says 星期, the answer is xīngqí, whatever the Standard is set to; when a Mainland voice says it, xīngqī: the reading you're marked on is the one the voice's clip says.

What Guoyu does not change yet: the tone judge compares you with Mainland voices. A properly calibrated Guoyu judge needs Taiwanese reference recordings and is planned. Characters are shown in simplified form for now; traditional display is planned too.

声调 Tones

Three modes, chosen from the first drop-down.

single syllables (tone dictation): hear one syllable, tap its tone. tone pairs: hear a two-syllable word, tap both tones on the grid (the fifth column, 轻, is the neutral tone).
repeat back: hear it, then ● Record and say it back. Recording stops by itself after about two seconds. You then see a graph and, for a single syllable, a verdict (below).

The second drop-down chooses where items come from: HSK 1, HSK 1–3 or any syllable (a random syllable in one of the tones Mandarin words actually use). Combinations that no Mandarin word uses, such as ché or kěng, are never drilled. Words ending in 儿 (哪儿, 一点儿 …) are left out of the tone drills: erhua is one syllable, and it sounds unnatural spoken alone. With "library voices for single syllables" ticked, a one-syllable HSK word is played as its syllable from the tone library; unticked, as the voice reads the word. The voices take turns (Settings → Voices), because hearing many voices is what makes the skill transfer.

The voice library. Every syllable the tab drills, in all four tones, is a stored clip in four Microsoft neural voices (Xiaoxiao, Xiaoyi and Xiaohan for Mainland, and 曉臻 HsiaoChen for Taiwan), and every HSK 1–3 word in seven (with Yunxi, Yunyang and Yunjian), made once. Every tone clip was measured before it was published. The male voices didn't keep their tones reliably, so they read words only, and the few syllables whose clip in one voice didn't keep its tone are played in another voice. The Mainland voices' third tones were shallow, so on 23 October 2026 those that didn't dip clearly were made again, deeper (730 syllable clips and 759 word clips), and zhao and lian, with 找 and 脸, play in their own voices again; a device that kept the old clips takes the new ones as it plays them. Everything here plays from that library: nothing is made on the spot.

After you answer, the item is revealed with its characters, its readings in your chosen form, and its meaning.
Show meaning (since v0.32.1), beside 🔊 Hear it, is a hint before you answer: a word's meaning, or for a syllable on its own (any syllable) up to three HSK 1–3 words with that sound and tone, a character alone before one inside a longer word (码, as in 号码), without their readings until you answer. A tone that no word in HSK 1–3 has shows a word for it from the whole HSK list instead, with its level and not a tap (since v0.32.2: 颠, as in 颠覆 (HSK 7–9)), or says it has none; a tap on any other character or word opens it to write. In repeat back the meaning comes with the item, as its reading does.
四声 four tones plays the syllable in each of its tones in one voice, only the tones Mandarin words use (for se that is just sè), labelled in your chosen readings. Use it after a wrong answer. Since v0.32.1 a tone's words show below the four when you point at it (or move to it with Tab); tap it (or press Enter) to hear it again.
manual next / auto next: whether the next item comes by itself 1.5 s after a right answer. Pressing "Hear it" cancels an automatic advance.
🐢 slower (ticked at first, since v0.30.1): every syllable and word, and four tones, is played a little slower, at 0.8× of the voice's speed, the pitch kept; Settings → Voices → Speed sets how much (0.5× to 0.95×). Untick it to hear the voices at their own speed.
Prepare 40 keeps the next items on the device so rounds work offline. For a whole voice, use Settings → Voices → Prepare this voice.
✍ then trace the characters: after a round from the HSK lists, a writing sheet opens with the round's characters, the ones you got wrong first, to trace once each, then offers Once more from memory. See 写 Write.

Repeat back: the graph and the verdict

Grey is the voice's pitch; teal is yours. Both are drawn relative to each speaker's own voice, so a low male voice and a high female one compare fairly.

Single syllables. heard as is the lab's judgement of which tone you produced. Each of the syllable's four tones in a Mainland voice gets a score from your contour and its pitch level, and (since v0.28) from your syllable's length and creakiness compared with what native speakers do in each tone: creak counts only where the tone wouldn't have it, and length only outside the tone's usual range, measured against your own average once you've recorded a few. The tone that scores best is heard, but the tone you were asked for is heard if its score is near enough the best (since v0.33.1), except that a contour rising like a second tone without dipping first isn't taken for a third. If the recording lost your pitch before the syllable ended, you're compared with the same first part of the voice's. similarity says how close your shape was to the target (for a third, the nearest of its forms: the voice's own, its low part, or a full fall-and-rise): 100 is identical, 80 is about half a semitone off on average, 50 a semitone and a half, and 0 three semitones or more (since v0.33.1; it used to reach 0 only at six, so the Log's earlier rounds show higher figures). A ✓ with a low similarity means "right tone, rough shape"; a ✗ with a high one, a shape near the target's but nearer another tone's, or said in a way the tone isn't (too high or low for you, too long or short, or creaky): the reason says which. After a ✗ the verdict says why: the tone your shape was nearest, or, when another tone was heard, what tipped it, compared with the tone your shape was nearest ("shape: nearest a fourth tone (72%), but longer and higher than a fourth tone for you, as a first tone is").

Words. A two-syllable word gets a guide instead of a verdict: your contour for the whole word drawn over the voice's, and how similar they are. Tones inside words, especially the last syllable, which is often short and creaky, can't be judged reliably by pitch alone from a phone recording, so the lab doesn't pretend to.

The third tone. Said alone, a full third falls and rises again. Since v0.33.1 the judge compares a third with the voice's own third and with a full fall-and-rise (the typical shape of six native speakers' thirds), so a full third counts even where the voice's own barely rises. In speech the third tone is usually just low (often creaky) without the final rise, and in Taiwan it is usually said that way even alone: the judge accepts that as a third tone when it's low for you or creaky, with its lowest point within its first 70% (or later, where the voice's own third is lowest later too), and says "low third"; when a contour shaped like one isn't low for you or creaky, the verdict says that's what it lacks. A contour that falls all the way to its end is a fall, not a third, unless the voice's own third does too, as many of Xiaohan's do.

Your recordings are never kept. Each one is judged on your device, then thrown away: it isn't uploaded, isn't kept on the device, and isn't in a backup.

How reliable is it? Measured on 30 October 2026 (v0.33.1) against the library's three Mainland voices, the ones the judge compares you with. On three native speakers' recordings made on their own phones, the judge accepted 37 to 40 of 43 correct single syllables, depending on which voice it compared with (28 to 31 of the 32 from the two Mainland speakers; before v0.33.1, 32 to 33, and 24 to 26). It let through none of their 33 deliberately wrong tones (a fourth tone said for a first, second or third, and a first for a fourth), but that's a small sample, and a mistake can pass: checked against a tone they weren't, their correct tones were accepted about 1 time in 10, most often a second as a third (9 times in 15) and a fourth as a third (4 times in 27). On a published collection recorded by six native speakers (2,733 to 2,847 syllables, depending on the voice) it heard the right tone 86 to 88% of the time when it wasn't told which tone was meant, up from 82 to 83%, and accepted 90 to 92% when it was told, as it is when you practise; third tones were the hardest (74 to 86% of them accepted), and their thirds were accepted as seconds 4 times in 10 when a second was asked for. Neither test is independent: the six speakers' recordings set the judge's allowances for creak and length and its full third, and the three speakers' were used to choose its other settings, so on other people's voices expect it to do a little worse. The samples of phone recordings are small, so treat these figures as indicative, and a single ✗ as a prompt to listen again rather than a ruling. If a recording is very quiet, the verdict says so: hold the phone a little closer, in a quiet room.

Spoken tones and written tones

Dictionaries write 你好 as 3·3, but everyone says 2·3: when two third tones meet, the first becomes second. 一 and 不 change too (一起 is said 4·3, 一定 2·4, 不是 2·4), but 一 keeps its first tone in 第一 and in dates (一月). Readings show the dictionary form; the pair quiz and repeat-back expect the spoken form. Where they differ, both show the word as it's said first (since v0.30.1: 不但 búdàn), then as it's written and why (written bùdàn: 不 before a fourth tone is said in the second tone).

Six words the word list gives both ways are accepted both ways (since v0.31): 里面 lǐmiàn or lǐmian, with a light 面, and 前面, 上面, 外面 and 下面 the same; 知道 zhīdào or zhīdao. The pair quiz and Dictation accept either; the pair quiz's reveal and repeat-back show both.

Putonghua and Guoyu. Taiwan Guoyu uses the same four tones as Putonghua but differs in places: fewer neutral tones (东西 dōngxī, 谢谢 xièxiè), the low third described above, a second tone that often starts with a slight dip, and a few words with other tones (星期 xīngqí, 垃圾 lèsè). Both are correct Mandarin; choose yours under Standard.

听写 Dictation

A word from the HSK list plays; type what you hear, in pinyin or in 注音 (choose type pinyin or type 注音). The first drop-down chooses the level: HSK 1, HSK 2 or HSK 3 words, or all of HSK 1–3. Press Enter (or check ↵ on the keypad) to check, and again for the next.

Pinyin, Skritter style: tone numbers, ni3 hao3; nv3 or nu:3 for nǚ; 5 (or 0) for the neutral tone. Spaces are optional; erhua can be dian3r or dianr3.
注音: on a phone, a zhuyin keypad replaces the keyboard; on a computer the keys type zhuyin in the standard Taiwanese 大千 layout (1 = ㄅ, q = ㄆ, a = ㄇ …; 6 3 4 7 for ˊ ˇ ˋ ˙; space or ˉ for first tone); a zhuyin input method works too. A tone mark ends each syllable, ˙ may go before or after, and an untoned ㄦ after a syllable is erhua.
Green: right sound and tone. Amber: right sound, wrong tone (the expected tone is shown). Red: wrong sound. Grey: a syllable you missed. Feedback is shown in the system you typed in.
🐢 Slower, 汉字 and English are hints you can use before checking. A round starts slower while 🐢 syllables and words slower is ticked (since v0.30.1): the same clip played more slowly, the pitch kept, at the speed Settings → Voices sets; 🐇 Normal plays it at the voice's own speed.
Starts with (since v0.32.2) is a hint too: each syllable's start, without its tone, on a line under the box: its first letter or two in pinyin (朋友: p· y·; 吃饭: ch· f·), or its first symbol when you type 注音 (朋友: ㄆ· ㄧ·).
Dictation expects dictionary tones (3·3 for 你好), because that is what you type into Skritter. Under Guoyu, a full tone is also accepted where Putonghua has a neutral one.
✍ then write the words: at the end of the round, a writing sheet opens with its words, prompted by their sound, then offers Once more from memory. Words you hear wrong also come back in Write today.

写 Write

Learn to write the characters you've been hearing. Each character is written on a square, stroke by stroke, and every stroke is checked as you draw it: its shape, its direction and its place in the order. ▶ Start round gives you about 12 characters, in whole words.

Source: Write today (below), HSK 1 or HSK 1–3. Cue, what prompts you: the word's sound (the default; 🔊 plays it again), its pinyin (or zhuyin, as you read), its meaning, or a mix of the three in turn. One syllable can be many characters, so under Sound the meaning is shown too for one-character words and for words that sound the same as another HSK 1–3 word; the meaning button is always there. When a word's sound can't be had just now (offline, before its voice is prepared), you get its pinyin instead.
Mode: Learn plays the strokes once, then you trace them over the outline, then write the character once from memory. Recall is from memory, with no outline. On paper: write it on paper, press Show to watch it played, then mark yourself ✓ or ✗; it suits a computer without a touch screen. Auto (the default) chooses for each word: Learn if one of its characters has never been written here, Recall otherwise.
Hint lights the next stroke; Show plays the whole character, then you trace it; Skip moves on. A stroke refused three times also shows you the next one. A character written with none of these is clean, even if you corrected a slip or two yourself; the mark under each character (✓, hint, shown, skipped) says how it went.
Write today brings back, in this order: the characters that needed help on an earlier day (learning them, a hint, Show, Skip or ✗); the words you heard wrong in Dictation; then, with add new characters ticked, new characters from the HSK list, HSK 1 first. A character is settled when you write it cleanly from memory (or get it right on paper) on a later day; until then it comes back each day. At the end of every round, the words you learned or needed help with come back once more, from memory, by sound.
Strictness sets how closely a stroke must follow the model: relaxed (the default on phones and tablets), normal (the default elsewhere) or strict. Tracing over the outline is more forgiving than writing from memory, and a stroke drawn out of order or the wrong way round is refused either way. Pen: on auto, once a pen comes near the square the lab ignores fingers and palms there (it says so), so your hand can rest on the screen while you write; pen only does that from the start; any takes fingers too, for writing with a finger. Palm-sized touches are always ignored.
From the other tabs: tick ✍ then trace the characters in Tones, or ✍ then write the words in Dictation, and a writing sheet opens at the end of the round. Each ends with Once more from memory.
Prepare 40 loads the stroke data, and under the Sound cue the voices, for the next 40 characters, so rounds work offline. Offline, a round uses only words whose stroke data is on the device.
Writing stats: characters written, how often you wrote from memory without help in the last 30 days, what's due today, and the characters you keep missing (tap one to write it). Copy my writing list puts the words you've written on the clipboard, weakest first, one per line, for a flashcard or writing app.
Where a character has a single traditional form that differs, it's shown small beside it once you've written it (學 beside 学).
Your strokes are checked in your browser and never sent anywhere.

记录 Log

What's prepared for offline use (the voices' clips, and the stroke data for writing), and every tone, repeat-back, dictation and writing round with its score; repeat-back rounds of words show their average similarity, as a guide. Per-tone accuracy and your most common confusions are under Tone stats in the Tones tab.

设置 Settings

This site: how to install the tone lab on your device, and Show the welcome card again.
Readings: Readings, Standard and tone colours (see Readings and standard).
Voices: the voice for Tones, Dictation and Write: Rotate · Mainland voices (the default for Putonghua), Rotate · Taiwan voices (the default for Guoyu), or one voice by name. Prepare this voice keeps all of one voice's tones and HSK words on the device (about 34 MB) for practising offline; a Taiwan voice brings its Mainland partner's tones for the judge (since v0.31, for a syllable the partner lacks, the four tones of the voice that judges it). It can be stopped, and picks up where it stopped. 🐢 syllables and words slower (ticked at first, since v0.30.1, in place of Slower replays) and its Speed (0.8× unless you choose): Tones and Write play everything at that speed, and Dictation's rounds start at it. The voices are Microsoft Azure neural voices, not recordings of people. The tone library and the HSK words were synthesised once and are stored by TingliLab.
Microphone: a level meter and a record-and-play-back test. Reset audio does what a page refresh does when the mic stops responding.
Data: Download backup saves your practice log and settings as a small file, without audio; Restore from file merges one in, here or on another device.
New versions (since v0.33.1): when a new version is out, a bar at the top asks you to reload; your practice is kept. To know, the page asks the server which version is live while it's on your screen: every half hour, when you come back to it, and when the connection returns (see Privacy). It waits for the end of a round.

Offline

The tone lab keeps itself on your device the first time it opens. The voices' clips are fetched as you use them and kept; Settings → Voices → Prepare this voice fetches all of one voice at once (about 34 MB: check you're on Wi-Fi), and Prepare 40 in Tones and Write prepares the next few rounds. Then tone practice, four tones, the judge, dictation and writing all work with no connection, and the welcome card and this page open too. Offline, rounds draw only on what is already on the device.

Privacy and data

The full statement is on its own page: how the tone lab handles your data. In brief: there are no accounts; your practice log, settings and backups stay on your device; recordings are judged on your device and thrown away; nothing is sent to our server but the ordinary requests that load the page and its audio, and now and then one asking which version is live, whose standard access logs are kept 15 days; no cookies, analytics or third-party scripts.

FAQ

Something plays with no sound / the mic stops working. Settings → Microphone → Reset audio, or reload the page.
"Offline, and nothing is cached". The voices' clips need the network once; afterwards they are kept on the device. Use Settings → Voices → Prepare this voice, or Prepare 40, before going offline.
The voices are too fast, or too slow. Since v0.30.1 Tones, Dictation and Write play a little slower at first (0.8×). Settings → Voices → Speed sets how much, and 🐢 in Tones turns it off.
I learned with zhuyin, not pinyin. Settings → Readings → 注音 (or both), and in Dictation choose type 注音.
I'm learning Taiwanese Mandarin. Settings → Readings → Standard → Guoyu. Read what it does and doesn't change yet under Readings and standard.
Can I see or write traditional characters? Not yet; characters are shown simplified for now. Where a character has a single traditional form, Write shows it beside the character once you've written it.
Write refused a stroke I drew right. Order and direction count: a stroke drawn out of turn, or the wrong way round, is refused. The stroke data follows the Mainland standard order, which differs from what Taiwan teaches for a few characters.
Why does a word get no ✓ or ✗? Because tones inside words can't be judged reliably from a phone recording by pitch alone; see Repeat back. Single syllables are judged.
My third tone keeps being marked wrong. In isolation the Mainland third tone dips low (often creaky) before it rises, and it is the longest tone; a low third without the rise is also accepted. Play four tones on the item and copy the third.
Is my progress saved? In this browser, on this device. Download a backup (Settings → Data) to keep a copy, or to move to another device.
Can I take my practice log to the full lab? Yes: download a backup here, sign in at the full lab, and use Settings → Data → Restore from file there.
Why is it free? Everything here plays from voices made once and stored, and runs in your browser, so it costs almost nothing to run, however much you use it. The full lab writes new listening material for you and talks with you, which costs per use (see The full lab).
Questions or feedback? Email hello@tinglilab.com.

The full lab

This is the free part of a bigger lab. The full TingliLab — listening clips written from your own word list, a conversation partner, texts you paste in, sync across devices, classes for teachers — is explained at tinglilab.com. Free accounts with daily limits are open now (100 places to start, 18 or over). A paid version with a larger allowance may follow if there is demand — if it does, it will feature no subscriptions, no auto-renewals, no card details stored, you renew when you like — but no price is set and nothing available there requires payment at this time. If paid accounts do happen, anyone with a free account created by 31 December 2026 will pay half the regular price for as long as they keep renewing.

About the full lab

Teachers and tutors: would you like to use this with your students? A limited number of teacher trial places are open now: write to hello@tinglilab.com with "teacher" in the subject. Students can use the free tone lab without accounts.

Credits

Founding Testers: Bangle Wu (Bernice, 吴帮乐), Kanghong Shi (史康宏), Yongfu (Deven) Huang.

Voices: Microsoft Azure neural text to speech (prebuilt voices), used under the Microsoft Product Terms; the stored libraries were made on the paid tier. Taiwan readings: checked against the Ministry of Education's Revised Mandarin Dictionary (教育部《重編國語辭典修訂本》). Word lists: HSK 3.0 (2025 revision), from complete-hsk-vocabulary (MIT), with meanings from CC-CEDICT (CC BY-SA 4.0; the licences: lib/LICENSES-2.txt). Since v0.32.2 the whole list, levels 1 to 7–9, gives a word for a sound and tone that HSK 1–3's words don't show. Pinyin: pinyin-pro. Stroke order and checking: Hanzi Writer (MIT), with stroke data from Make Me a Hanzi (Arphic Public License). Traditional forms: OpenCC (Apache 2.0). Countries in the visitor counts: IP Geolocation by DB-IP. Thanks to the native-speaker testers whose recordings and feedback calibrated the tone judge.

Changes

The tone lab shares its code, and this list, with the full lab; earlier changes are on the full lab's help page.

Version	What changed
v0.33.1	The tone judge: the tone you were asked for is heard when its score is near enough the best (not a third, though, for a contour that rises without dipping); a full fall-and-rise counts as a third whatever the voice's own third is like, and a third without its rise only when it's low for you or creaky; a contour that falls to its end is a fall, unless the voice's own third does too; your pitch level counts half as much as before; after a ✗ the verdict says why; similarity is on a stricter scale. When a new version is out, a bar at the top asks you to reload. Privacy statement, version 3 (the version check).
v0.33	A first visit downloads a little less: the tone lab no longer fetches the speech SDK, which only the full lab uses.
v0.32.2	In Tones, a sound and tone with no word in HSK 1–3 shows one from the whole HSK list, with its level (颠, as in 颠覆 (HSK 7–9)). Starts with in dictation, a hint: each syllable's start (its first letter or two, or 注音's first symbol), without tones. The panel about the full lab links to tinglilab.com.
v0.32.1	Show meaning in Tones, a hint before you answer: a word's meaning, or a syllable's words in HSK 1–3; each of the four tones' words on hover or tap. Characters such as 三 and 马, which the HSK list also gives as surnames, now come up in tone practice and dictation, and their meanings leave the surname out.
v0.32	Nothing changes here. The full lab's accounts gained a review schedule, their own word lists and a photo of a page into Texts.
v0.31	Six words the word list gives with a light syllable too are accepted both ways: 里面 lǐmian (and 前面, 上面, 外面, 下面), 知道 zhīdao. Tone stats and the Log's repeat-back rounds say matched where they said right. A remade clip slow to come plays the copy on the device after 3 seconds, the new one kept for the next play. Prepare this voice for HsiaoChen also brings the four tones of the voice that judges a syllable her partner lacks. The lab's icon on the tabs of help and privacy.
v0.30.1	Tones and Dictation: single syllables and words a little slower at first (0.8×, the pitch kept; Settings → Voices sets the speed, in place of Slower replays, and 🐢 in Tones turns it off). Tones' reveal and repeat-back show a word as it's said first (不但 búdàn), then as it's written and why. Two syllables whose third tone sounded like a second are played in other voices (zhao from Xiaoxiao and Xiaoyi, lian from Xiaoxiao), and so are the words 找 and 脸, even on a device that kept them. Write: its words at Tones' speed too, and a meaning's Chinese characters a little larger again.
v0.30	Write: in a word's two-line meaning, its Chinese characters are a step larger (the English as it was). Questions and feedback now go to hello@tinglilab.com; the tone lab still has no accounts and sends no email. Privacy statement, version 2 (the address).
v0.29	Meanings in full (they stopped at 70 characters): in Tones, Dictation and Write a long one takes up to two lines and ends at a whole sense with "…", with the rest on hover; a few slang and vulgar senses are left out here, and notes that only point to another way of writing a character are gone. The full lab's free places, under The full lab and at the foot of Tones, show as they are: open, or taken for now.
v0.28	The tone judge takes a syllable's length and creakiness from what native speakers do in each tone (the library's voices carry neither), the length against your own average, and compares a syllable whose pitch was lost early with the same first part of the voice's: correct native tones accepted rose from 76–83% to 84–90% on a set of 4,100 syllables, and from 24–29 to 34–38 of the testers' 43, with every wrong tone still caught. A tone play the item's voice lacks comes from another library voice, and a voice withdrawn from the tones keeps its words.
v0.27	The free tone lab opens at tinglilab.org: tone practice with the judge, dictation on the HSK 1–3 words by level, and writing, with no account and no limits, on the voice library alone; nothing is kept on a server and nothing is synthesised. Slower replays play the stored clip more slowly. Its own help and privacy pages. On both sites, the four-tones play and the judge always use a voice whose four tones of the syllable are in the library.
v0.26	The tone and word libraries. Every syllable the Tones tab drills, in all four tones, is now a stored clip in four Microsoft neural voices (three Mainland, one Taiwan), and every word on the HSK 1–3 list in seven (with three male Mainland voices, whose tones measured too unreliable for tone practice), made once and served as files: tone practice, the tone judge, the pair quiz, and dictation and writing on HSK words no longer use the speech service, and work offline after Settings → Voices → Prepare this voice. One voice choice for Tones, Dictation and Write. Each item is marked by the standard of the voice that said it, and the Taiwan readings were checked against Taiwan's Ministry of Education dictionary. The voices are synthetic and say so.