Why AI Singers Mispronounce Chinese Lyrics, and How to Fix It
Characters like 长, 乐 and 行 trip up AI singers. What we measured with Google's Lyria 3.5, why pinyin notes don't help, and the homophone and kana swaps that do.
By the NeonSong team · Published
Write a Mandarin song about growing up, and an AI singer may sing 长大 ("to grow up", zhǎng dà) as cháng dà, which isn't a word. Write about music, 音乐 (yīn yuè), and you might hear yīn lè. Listeners notice at once, and the song is spoiled by one syllable.
This post explains why it happens, what we measured when we tested Google's Lyria 3.5 model in October 2026, and the fix that works for Chinese and for Japanese: change how the word is spelled for the model, not how it's annotated.
The problem: one character, several readings
Many Chinese characters (多音字, duōyīnzì) are read differently depending on the word they're in:
| Character | In this word | Read | But usually | As in |
|---|---|---|---|---|
| 长 | 长大 (grow up) | zhǎng | cháng | 很长 (very long) |
| 乐 | 音乐 (music) | yuè | lè | 快乐 (happy) |
| 行 | 银行 (bank) | háng | xíng | 不行 (no way) |
| 重 | 重新 (again) | chóng | zhòng | 重要 (important) |
| 朝 | 朝着 (towards) | cháo | zhāo | 朝阳 (morning sun) |
| 调 | 调皮 (naughty) | tiáo | diào | 声调 (tone) |
A human singer knows which reading the word needs. A singing model has to pick a reading for each character from what it learned, and when the right reading is the less common one, it often picks the common one.
Japanese has the same problem with kanji, which can have several readings each. In our test, 生の声 ("a live voice", nama no koe) came back with 生 read as sei, and 二人の間 ("between the two of us", futari no aida) with 間 read as ma.
What we measured
In October 2026 we generated test songs with Lyria 3.5 from original lyrics containing these words, separated the vocals, and ran them through speech recognition plus two automatic pronunciation judges. Some honest caveats first: the samples are small, and the judges themselves disagreed with the expected reading about 30% of the time when we fed them unambiguous text, so read the numbers as trends, not verdicts.
Mandarin, written normally. The tricky words were sung with the intended reading only about a third of the time (36% of the words the judges could decide).
Mandarin with pinyin notes. Adding the reading in brackets, like 长(zhǎng)大, made no difference. The model obeyed "don't sing these notes" and simply ignored them: 长大 was still cháng dà.
Mandarin with homophones. Replacing the risky character with a different character that has the same sound and only one reading, so 长大 becomes 掌大 (掌 is always zhǎng), raised the intended reading to about two thirds (67%). 音乐, 银行, 朝着 and 调皮 improved the most. It isn't a guarantee: in a retest, 掌大 was still heard as cháng in two of four takes.
Japanese with kana. Writing the tricky words in kana instead of kanji (なまの声, ふたりのあいだ) brought the error rate down to 6–10%, close to English, and none of the test words was misread. Reading notes didn't help here either.
Cantonese. Lyria does sing Cantonese rather than Mandarin when asked: Cantonese-only words like 嘅, 唔 and 佢 come through. How accurate the tones are needs a native listener, which an automatic test can't replace.
How to fix it yourself
These tricks work with any AI music tool that takes lyrics as text:
- Find the risky words. Look for characters from the table above, and any word where the reading you want isn't the character's most common one.
- Swap in an unambiguous homophone. Pick a character with the same sound and only one reading: 长大 → 掌大, 音乐 → 音月, 银行 → 银航, 重新 → 虫新, 朝着 → 潮着, 调皮 → 条皮. The model only needs the sound; the meaning of the stand-in doesn't matter.
- For Japanese, write the word in kana. 生の声 → なまの声. 間 → あいだ.
- Keep the original lyrics for display. The swapped version is only for the model. Use your original text in captions, lyric videos and the song description.
- Listen to the result. Swaps raise the odds; they don't guarantee the take. Generate two versions and keep the one that sings it right.
How NeonSong does this for you
NeonSong's Reading Check does those steps automatically before you spend credits. It flags lines with characters that are likely to be misread, suggests the homophone or kana spelling the model sings correctly, and lets you accept or change each suggestion; the swaps you make are remembered for your next songs. Your song page always shows the lyrics exactly as you wrote them.
Where the right reading can't be worked out from a dictionary, Reading Check says the reading is a guess instead of swapping silently, so a wrong guess never makes a song worse.
Try it with your own Chinese or Japanese lyrics. New accounts start with enough free credits for a few songs.
About the test
NeonSong's internal benchmark of Google's Lyria 3.5, run on October 22, 2026 with a follow-up on October 23 after Reading Check shipped. Lyrics were written for the test. Vocals were separated with Demucs before recognition; Mandarin pronunciation was judged by a CTC phoneme comparison and a pinyin match, and Japanese by kana recognition. Error rates are character error rates of the recognized vocals (English reference: 5%).