Turn your lyrics into a song with AI
Write it is the mode in Midhull for people who already have the words. Paste or type your lyrics, tag the sections, describe the sound in one line, and you get a full song with your words sung as written, in the language you wrote them in. This page covers how to format lyrics so an AI singer gets them right, what we measured in our detailed test of the first seven languages, and what to do when one line comes out wrong.
Updated

If you are searching for a song from lyrics you half remember, that is a different job: a lyrics search engine will find it. This page is about turning words you wrote into music.
What happens to your lyrics
- Your lyrics are read as a structure. Each tagged block ([Verse], [Chorus], [Bridge]) becomes a section of the song, in the order you wrote them.
- The sung language follows your words. Lyrics written in Korean, Japanese or Hindi script are recognised as such; otherwise the song follows the language you ask for or your app language, and you can set it by hand.
- Your one line description sets the sound. Genre, tempo, instruments and voice come from the description; the words come only from you.
- You get two takes. Each is a full song with your lyrics. Keep the better one.
- You fix lines, not songs. If one word is sung badly, re sing that line. If you want to change a word, rewrite that line. Lines you are happy with can be locked so they do not change.

How to format lyrics for an AI singer
Most "it ignored my lyrics" complaints come down to formatting. The engine needs to know where sections start and how the lines group.
| Do | Why |
|---|---|
| Put a tag on its own line before each section: [Verse], [Pre-Chorus], [Chorus], [Bridge], [Outro] | The engine maps each block to a part of the song |
| Keep 4 lines per verse and chorus to start | Four line sections fit common 8 and 16 bar phrases |
| Keep line lengths within 2 or 3 syllables of each other inside a section | Even lines get an even melody; one long line gets rushed |
| Repeat the chorus word for word each time | It tells the engine this is the hook |
| Write numbers and abbreviations as words ("twenty one", "doctor") | You control how they are sung |
| Spell an unusual name the way it sounds ("Shivon" for Siobhan) | The singer reads what you wrote |
| Leave out stage directions inside lines | Text inside a line will be sung |
Supported section tags: intro, verse, pre chorus, chorus, post chorus, bridge, rap, drop, breakdown, solo, interlude and outro.
Syllables decide the rhythm
Count the syllables in each line before you send. Here is a verse we wrote for a German piano ballad, "Komm nach Haus", with the count per line:
[Verse]
Ich hab dein Lachen noch im Ohr, (8)
wie Regen leise an der Scheibe. (9)
Die Stadt ist stiller als zuvor, (8)
und ich weiß nicht, wo ich bleibe. (8)
Eight, nine, eight, eight, with rhymes on lines 1 and 3 (Ohr, zuvor) and 2 and 4 (Scheibe, bleibe). That evenness is why it came back near perfect: a word error rate of 0.01 when we checked the vocal with speech recognition.
Listen
Sunrise, take the long way
Summer pop · Bright
1:30
Lyrics
Windows down and the radio on
Coffee's cold but the night is gone
Five of us in a borrowed car
Chasing light to wherever we are
Sunrise, take the long way home
Sunrise, we were never alone
Hold the moment, let it glow
Sunrise, sunrise, don't let go
Sunrise, take the long way home
Sunrise, we were never alone
Hold the moment, let it glow
Sunrise, sunrise, don't let go
Komm nach Haus
Piano ballad · Tender
1:30
Lyrics
Ich hab dein Lachen noch im Ohr,
wie Regen leise an der Scheibe.
Die Stadt ist stiller als zuvor,
und ich weiß nicht, wo ich bleibe.
Komm nach Haus, das Licht ist an,
ich hab den Tisch für zwei gedeckt.
Komm nach Haus, so schnell du kannst,
bevor der Winter uns entdeckt.
Komm nach Haus, das Licht ist an,
ich hab den Tisch für zwei gedeckt.
Komm nach Haus, so schnell du kannst,
bevor der Winter uns entdeckt.
Baila conmigo en la cocina
Latin pop · Playful
1:30
Lyrics
Domingo lento, café en la mesa,
suena la radio, se va la pereza.
Tú con mi camisa, yo sin zapatos,
bailando torpes entre los platos.
Baila conmigo en la cocina,
que la mañana se vuelve divina.
Gira, gira, que nadie nos mira,
y el mundo entero se para y suspira.
Baila conmigo en la cocina,
que la mañana se vuelve divina.
Gira, gira, que nadie nos mira,
y el mundo entero se para y suspira.
천천히 와도 괜찮아
R&B · Warm
1:30
Lyrics
비 오는 창가에 기대어
네가 남긴 향기를 맡아
식은 커피 한 잔 사이로
너의 목소리가 들려와
천천히 와도 괜찮아
나 여기서 기다릴게
밤이 길어도 괜찮아
네가 올 때까지
천천히 와도 괜찮아
나 여기서 기다릴게
밤이 길어도 괜찮아
네가 올 때까지
まだ帰らないで
City pop R&B · Late night
1:30
Lyrics
改札を抜けたら
夜風が少し甘い
君からのメッセージ
既読のままでいい
ねえ まだ帰らないで
この街の灯りが消えるまで
ねえ ゆっくり歩こう
二人だけのリズムで
ねえ まだ帰らないで
この街の灯りが消えるまで
ねえ ゆっくり歩こう
二人だけのリズムで
लौट के आऊँगा
Folk · Longing
1:30
Lyrics
गाँव की पगडंडी, मिट्टी की खुशबू,
नीम की छाँव में बैठी है आरज़ू।
माँ की रसोई से उठता धुआँ,
याद दिलाता है अपना जहाँ।
लौट के आऊँगा, मेरे गाँव,
तेरी गलियों में रखूँगा पाँव।
शहर की रौनक भी फीकी लगे,
जब तेरी याद के दीपक जगे।
लौट के आऊँगा, मेरे गाँव,
तेरी गलियों में रखूँगा पाँव।
शहर की रौनक भी फीकी लगे,
जब तेरी याद के दीपक जगे।
Measured results for the first seven languages
Midhull sings in 37 languages, each one tested before we added it. The first seven are the ones we measured in detail: we sent lyrics we wrote in each of them to Midhull's sound engine and checked every vocal with Whisper large v3, a speech recognition model, comparing what it heard with what we wrote. Lower is better. Word error rate (WER) is used for languages written with spaces; character error rate (CER) for Japanese and Korean.
How to read it:
- German, English, Korean and Japanese came back with nearly every word where we put it. The Japanese differences are mostly kana and kanji spelling variants in the transcript rather than missing lines.
- Spanish sang every line correctly; most of its score comes from the engine adding an extra repeat of the chorus, which counts as inserted words.
- French and Hindi sang every line, with a higher error rate in the verses. For these two languages, plan to use the studio: re sing a line that comes out unclear, and keep the lines that are right.
- Speech recognition is a strict judge of singing: it also mishears real singers. Treat these numbers as a floor for comparing languages, not a quality score.
In the app


Lyrics we sent, with the songs they became
English, summer pop. Prompt: "Upbeat summer pop, bright synths and handclaps, female lead." Song: "Sunrise, take the long way" (WER 0.04).
[Chorus]
Sunrise, take the long way home
Sunrise, we were never alone
Hold the moment, let it glow
Sunrise, sunrise, don't let go
Spanish, Latin pop. Prompt: "Pop latino alegre con guitarra acústica y palmas, voz femenina." Song: "Baila conmigo en la cocina".
[Chorus]
Baila conmigo en la cocina,
que la mañana se vuelve divina.
Gira, gira, que nadie nos mira,
y el mundo entero se para y suspira.
Korean, R&B. Prompt: "한국 R&B, 미디엄 템포, 부드러운 남성 보컬, 기타와 신스 패드." Song: "천천히 와도 괜찮아" (CER 0.05).
[Chorus]
천천히 와도 괜찮아
나 여기서 기다릴게
밤이 길어도 괜찮아
네가 올 때까지
Japanese, city pop R&B. Prompt: "シティポップ寄りのR&B、ファンキーなベース、女性ボーカル、95BPM。" Song: "まだ帰らないで" (CER 0.13).
English, full length. Eight sections, 3:35, every section sung in order: "Porch light" (WER 0.06). If you can write a whole song, the engine will sing a whole song.
When a line comes out wrong
| Problem | Fix in the studio |
|---|---|
| One word mispronounced | Re sing that line; if it persists, respell the word the way it sounds |
| A line is rushed | Cut one or two syllables, or split it into two lines |
| The chorus does not land | Regenerate the chorus section only |
| The song ends before your last section | Extend it, or check the last section has a tag |
| The melody of the verse is fine but you want new words | Rewrite that line |
| Right words, wrong feel | Change the vocal style, the tempo or the key, or make it acoustic |
Studio tools open up as you make songs and level up: rewriting and re singing a line are there from your first song, key and tempo open at level 3, regenerating a section and adding a bridge at level 5, swapping an instrument at level 8 and stem export at level 10. Extending a song, changing the vocal style, making it acoustic and remixing into another genre are open from the start.






Frequently asked questions
Will the AI change my lyrics?
In Write it, the words you send are the words that get sung. Midhull's lyric writer only writes when you ask it to, for example in Describe it.
How long can my lyrics be?
A full song's worth. Our longest test, eight sections and 3:35 long, was sung in order. You can also extend a song afterwards, 10 to 90 seconds at a time.
Can I mix two languages in one song?
The sung language follows your lyrics or the language you choose, and you can set it by hand. Mixed language songs are not part of our language testing, so treat them as an experiment: set the main language, keep the other language to a short repeated phrase and expect to try more than once.
Can it sing in my voice?
Yes, after you verify your voice with a live consent read. See AI song with your own voice.
Can I download the song with my lyrics?
Yes, with a plan: MP3 or WAV, a chord MIDI file and the lyric sheet, plus stems once you reach level 10. Downloads of songs you make are never locked later. Which plans include stems and MIDI is confirmed with the plans at launch.
Is it free?
Your first full song is free during setup. After that, songs come with a plan; prices are published at launch.
Related
- AI song generator
- Need words first: AI lyrics generator
- The full method: How to make a song with AI
- Songs beyond English: AI songs in other languages
- Rap verses: AI rap generator
- Compare apps: Best AI music generator apps
- In the app: Create, Studio, Languages