About

We started with a song we couldn’t fix.

Ines had spent ten years in a publisher’s A&R team when the first AI song generators arrived. She tried them on a chorus she had been stuck on for months. Every result came back finished and wrong in one place, with no way to change just that place. In 2024 she and Kwame built the opposite: a generator that writes a song one track at a time.

Headphones on a wooden desk beside a notebook of handwritten lyrics

Halfsung is that idea, grown up.

Today we are fourteen people in Lisbon and London: writers, producers, two researchers and one person whose only job is paying session musicians. We are funded by people who work in music, not by a fund that needs us to replace it.

We build for one moment: when an idea is real but the song isn’t there yet. Halfsung plays the parts you can’t, one lane at a time, and waits while you change your mind.

How we work

Four rules we don’t bend.

  1. 01 / No track is final until you say so

    Every track can be redone on its own, as often as you like, without touching the others. Lock what you like; nothing locked ever moves.

  2. 02 / We learn from people we pay

    Our models are trained on session recordings made for us by 340 players and singers, paid by the hour and credited by name. Nothing a writer uploads trains anything.

  3. 03 / Uploads leave our servers within a day

    We read a hummed or played sketch, keep the notes and delete the audio within 24 hours. Your songs live in your DAW, not with us.

  4. 04 / The song is yours

    Melody, words, parts and stems belong to the writer, without a share, a credit or a catch. A tool should not be on the split sheet.

How our training data is licensed

Some of the team

14people, Lisbon and London
340session players paid for training data
24 hlongest we keep an upload