Create · Sing · Produce · Enjoy

· voice changer
MB the studiolocal · no account

create your music

Genre
Voice
Voice from a recording
Seed

nine genres, and the studio stays inside them: a style picked here outranks the sentence below.

or give it the words yourself

[verse] and [chorus] on their own line start a section, the same way the singer reads a sheet. One line is sung per bar, so the sheet decides how many bars the song has — every line is sung, in order, once. Leave this empty and the studio writes the lyrics from the brief.

style options

Say what it is about, and anything you want fixed: genre, tempo, key, male or female voice, length, instruments. “read it” shows what the studio heard before anything is rendered.

advanced · the two agents
The transformer decides what the song is; the producer builds it, judges its own take against a bar it wrote down first, and revises what fails. Nothing runs until you press.

receipt

nothing measured yet

workflow

  1. no run yet

recent creations

reading exports/…

lyric & melody

—
Produce a song to see its sheet.
the trained voice asking this machine…
Tick “sing it in a trained voice” by the brief. The engine here is RVC, run as a worker; if it is not installed or has no trained weights, this panel says which of the two is missing instead of quietly rendering the usual singer.

what it heard

nothing read yet

voice from a recording

a recording in voices/ becomes a voice the singer can use — pitch, brightness, breathiness and range measured from it.

how it says a word

Every word is looked up in this order: what you taught it, then the studio dictionary, then the built-in English list, then the spelling rules. Click a word in the sheet to see which one answered.

studio dictionary

Bemba is written phonemically, so the studio reads it letter by letter instead of guessing at English. Two rules hold: the syllables are the word's own letters, cut and never respelled, and the “said as” is exactly the phones of those syllables, in order.

add a word

correct the word on the screen

A correction is checked against the rules before it is kept: a split that changes a letter, or a “said as” whose vowel count does not match the syllables, is refused with the reason rather than stored.

the six parts

01

ghost producer

An agent that renders the take, judges it against seven criteria, and revises the one thing that failed — then hands over a receipt instead of an opinion.

measured · 3 attempts per brief, each failure named

02

rhythmatic

Every genre carries its own kick, snare and hat cells, bass figure and chord progression. The rhythm is the genre's, not a default with a tempo on top.

measured · nine genres, nine arrangements

03

all music pipelines

Brief or lyrics or recording in; song, five stems, a cloned voice, a cleaned take and a filtered vocal out. One studio, one receipt, no dead ends.

measured · stems complete at 513.384 s, 4 stems, 86.4 MB each

04

maestry producer

The mastering stage sets the master peak and gain, then the receipt is read back off the written file — peak, RMS, length, channels — so “mastered” is a measurement.

measured · receipt read from the exported wav

05

beat locker

Notes land on a bar-and-step grid and the grid is locked to the file, so a sung syllable and a kick drum cannot drift apart between sections.

measured · beat lock asserted in the studio suite

06

sing · sung

Every written line is sung word for word, one note per syllable, with phones from the dictionary and stress from the prosody model. No hummed placeholders.

measured · 0 unsung lines at 16 and 64 bars

The names are the studio's own. The numbers above come from the suites in _test/, not from this page.

assistant

Tell me what to change — “slower”, “in D minor”, “male voice”, “2 minutes”, “add strings”, “no drums”, or “why”.

stems · separate a song

or upload a file to separate

A WAV from anywhere — a mix, a recording, someone else's song. It is copied into exports/uploads/ and then appears in the list above, because a separator needs a file the studio owns and a browser cannot hand it a path.

kits — the cloud separator

Kits runs the separation on their machines and returns vocals and backing. The base URL and path are Kits’ own documented values (https://arpeggi.io/api/kits, /v1/stem-splits), so the key is the only thing to supply. It is written to .env on this machine and never sent back to the page. A key is created at app.kits.ai/api-access.
A song this studio exported can be split into vocals, drums, bass and the rest. The job is queued, run by the Python Demucs worker, and its stems land in a private bucket that only the route below serves — the same four tables the schema migration describes, kept in stems/index.json because the studio has no database.