Drang Studio

Direct the song.
Don’t roll the dice.

Drang Studio is an AI music studio for musicians. You pick the melody, shape the parts and choose the singer, and it checks every sung word. We built it on the two hardest cases first: technical prog rock, and Persian. Easier songs come with them.

See what’s proven

Hear a take

The short vowels Persian leaves unwritten

تَپِشِ tapesh-e beat قَلبِ ghalb-e heart پُر اَز por az full of اُمیدِ omid-e hope مَن man my

My heartbeat, full of hope,

تَپِشِ tapesh-e beat زِندِگیه zendegi-ye of life

is the heartbeat of life.

0:00

One of the hard cases. Persian is written without its short vowels, so a model has to guess them, and it often sings the wrong word. Drang writes the vowels in, then checks every sung word.

Hardest first

We built and tested Drang on the two hardest cases we could find, so the easier songs come with them.

1

Technical prog rock

Odd meters, polymeter, and long songs with real structure: an intro, a solo, a chorus that comes back.

7/8, 5/8, 7/4

Each heard as written, 3 seeds each. Polymeter too, with the riff on both lines.

2

Persian

Short vowels aren’t written, so AI singers guess them and get words wrong. On a strict rock bar the base model passed 0 of 20 songs, mostly on vocal delivery.

4.14 / 5

Persian pronunciation with the vowels written in, rated by 26 native listeners (139 ratings).

And the easier case: the same singers also sing English.

Sources: V2_PLAN.md, docs/BASELINE_PROTOCOL.md, listener survey (8–12 Oct 2026)

How a song gets made

  1. 1

    Words

    Paste your lyric, in Persian or English. The studio checks it, splits it into parts, and can leave some parts to the band alone.

  2. 2

    Singer

    Pick a named singer. Every voice is your own, one you generated, or a singer who agreed.

  3. 3

    Sound

    Pick style chips. Your style is sent exactly as written.

  4. 4

    Melody

    4 tune sketches, or your own melody from audio, MIDI or a score. Move any part higher or lower.

    • Write it for me
    • Use my melody
    • Key change from here

  5. 5

    Make

    You see the cost before every click. Keep the takes you like.

    • I like the start: redo the rest

What’s proven, what’s being tested, what we’re working on

Every result has its number and the file it comes from. Each lit segment is a real render, take or clip.

Proven

  • 7 of 8

    Real song structure

    Intro, solo, and a chorus that comes back. Fixed wall-to-wall singing in 7 of 8 renders.

    studio/arrangement.py

  • 18 of 18

    A singer keeps their identity, in Persian and English

    A blind listener named 18 of 18 clips correctly. Each singer sang 3 Persian and 3 English songs.

    V2_PLAN.md

  • 3 of 3

    Words are checked against a real singer

    Kei passed 3 of 3 takes under the bar set by a human singer (0.157, 0.119, 0.188 vs 0.215).

    V2_PLAN.md

  • 7/8, 5/8, 7/4

    Odd meters

    12 renders (4/4 control, 7/8, 5/8, 7/4; 3 seeds each) were heard as written. The three 4/4 controls measured 109.1–111.1 BPM against 110.

    V2_PLAN.md

  • 7 over 4

    Polymeter

    With the riff on both lines, the measure read all 3 test takes as a 7-cycle and all 3 controls as 8. The listener heard it in 1 of 3. A riff on one line alone was ignored.

    V2_PLAN.md

  • +2

    Key change

    Exactly +2 semitones on both seeds.

    explore/README.md

  • 7 of 9

    Continue a take without a seam

    7 of 9 joins “no seam”, 2 “slight”, 0 “noticeable”.

    V2_PLAN.md

  • 0.93–0.99

    Reading a melody from audio

    Match on 10 known songs, after lining up tempo and start (raw score 0.51).

    docs/MELODY_CONTROL_PLAN.md

  • −14.0 LUFS

    Release-ready loudness

    True peak −1.4 to −2.9 dBTP. Stems add back to the take at 0.99.

    docs/STUDIO_CONTROL_PLAN.md

  • $0.15

    Honest cost

    About $0.15 per render, shown before every click, with a hard daily cap.

    V2_PLAN.md, engine/budget.py

Being tested

Built, and marked “try it” in the studio.

  • Fit the rhythm to the words
  • Time choice
  • Guitar plays the chorus
  • Other chords here
  • Best-of-N ranked by sung words
  • Humming in a melody

Working on

  • Duets

    Several singers in one song: duets, lead and harmony, a guest verse.

  • Mid-song shifts

    Changing tempo or time signature inside a song. Joining separate renders at a tempo change didn’t work (12 of 12 joins noticeable), so we’re trying a new way.

  • More singers

    More voices that pass the words check. Kei passes; two others don’t yet (0.276–0.528).

  • Small edits

    Changing one part without re-rendering the whole song.

  • Languages

    Languages beyond Persian and English.

Unless noted, each result was judged by one expert listener.

Singers you design, never cloned

“Real artists’ voices are never cloned.”

  • Your own voice

  • A voice you generated from your own prompt

  • A singer who agreed

Artist names are used only as style words.

Since January 2024

We started making music with AI in January 2024. We tested the existing models, made thousands of songs with them, and found an audience. Drang is the engine we built from what those models got wrong.

  • 10,000+

    Songs generated on existing AI music platforms.

  • Terrauhn

    Our artist name, with monthly listeners on Spotify and other platforms.

  • 38%

    Of native listeners’ answers said they would send the song to a friend (4 or 5 out of 5), knowing it was made with AI. The most common answer was 5.

Survey: 26 native Persian listeners found through Telegram, Instagram and X, 139 ratings of 20 base-model songs in 4 days (8–12 Oct 2026).

The rule we build by: “Failed features ship hidden.”

For mentors

What we need

The engine works. We want help turning it into a product musicians keep coming back to.

  1. Product-market fit: which musicians to serve first, and how to reach them.
  2. Service design for retention: what brings a musician back for the next song.