Sing your songs in your own voice
On Plus and Pro, you read your lyrics aloud and Songcode sings them on your melody, in your voice. Every account also gets singers made from real people’s recordings, in English and Japanese.
You can now have a song sung in your own voice. You don’t sing: you read the lines aloud, in the voice you’d use to read anything, and Songcode puts your words on the song’s melody. The tune is already in the code, so the only thing it needs from you is how you say the words.
Your voice is part of Plus and Pro.
How it works
Write or generate a song with a vocal, open the vocal stem’s menu and pick Use your voice. You get the stem’s lyric lines, one under the other. Tap Read and say the line. It stops by itself when you finish, shows the reading’s waveform, and you can play it back before you move to the next.
Read, don’t sing. If you sing the line, Songcode drops your tune and uses the stem’s notes. A plain reading at your normal pace works best, close to the mic, somewhere quiet. If a reading can’t be used, it says why (too quiet, cut short) and you read it again.
When the lines are read, tap Sing it in my voice. The stem now names your voice like any other sound:
stem vocal: voice.v3x9q2k7m {
line [e4/4 g4 a4/2 g4/4 e4 d4/2]
lyrics 2bars {
"hold the light for me"
}
in chorus {
harmony chords
}
}
From there it’s a vocal like any other. Change a note and your voice sings the new note. harmony chords stacks copies of your voice on the chord under the line, so a chorus can open up into a small choir of you. Lines you read carry over to your next song, so the words you use a lot only need reading once.
What happens to your reading
The recording never leaves your browser. While you read, your device listens and measures: where each sound of the line starts, and every 5 milliseconds the shape of your voice’s tone and how much of it is breath. Those numbers are all we get. The recording stays in your browser only so you can play it back, and is dropped when you close the sheet.
To sing a note, the engine takes the measurements for that syllable, holds the vowel for as long as the note lasts, keeps the consonants at the speed you said them, and plays it all back at the note’s pitch. That’s arithmetic on your measurements, in your browser. No model listens to your voice, learns it or makes anything from it, and we never use it to train one.
A song that uses your voice names it by a code, so anyone you share the song with hears it in your voice. Nobody else can read lines into it or use it in a song of their own. Delete your voice, at the foot of the same sheet, deletes every line at once.
Singers, for everyone
The same way of turning measurements back into a voice is behind Songcode’s new singers, and those are on every account. Each singer is one real person’s voice, made from recordings they gave away: read for LibriVox and donated to the public domain, or dedicated to the public domain through Mozilla Common Voice and the Open Home Foundation’s voice datasets. We cut those recordings into short pieces, one for every pair of sounds the person said in a row, and measured each piece the same way your browser measures your reading. A sung note is the pieces for its syllable, joined and played at the note’s pitch. To cut a recording we need to know where each sound in it starts, so a small classifier finds those times once, when a singer is made. It gives times and nothing else, and no model makes or changes any sound you hear. The licenses page credits every source.
There are ten of them, in English and Japanese, feminine and masculine:
- Juliana (English, feminine): clear words, bright and close
- Louise (English, feminine): the clearest words, even and warm
- Tanya (English, feminine): soft and light
- Evan (English, masculine): clear words, warm and plain
- Crispin (English, masculine): clear words, low and even
- Duncan (English, masculine): deep and calm, the words a little softer
- Mari (Japanese, feminine): soft and clear
- Maki (Japanese, feminine): bright and quick
- Hideki (Japanese, masculine): calm and low
- Kenji (Japanese, masculine): bright and plain
Turn vocals on and Songcode picks a singer for the song, in the language your prompt is written in or asks for. To choose one, open Vocals in the composer and pick from Voice. Japanese singers sing kana, one to a note.
The six voices Songcode had before (soprano, alto, tenor, baritone, robot and whisper) are still there, now called vocal synths. They’re synthesizers shaped like a voice, and their words come and go, so they work best as layers around a singer: a robot answering a line, a talk box under the chorus.
What to expect
Your voice and the singers are people reading, placed on a tune. The words come through and the voice sounds like the person, close to the mic. They don’t belt or run the way a trained singer does, and they’re at their best on melodies that move by step, with a syllable on every eighth note or longer. If you want a vocal that sounds like a finished record, a service that generates vocals will get you closer. If you want your own voice on it, read the lines.