Seed
3 November 2026 · Day 3 · Built with Claude, on 5 November
When a sung or played phrase becomes notes, what is kept and what is lost? A recording and what a machine made of it, side by side.
The recording
The waveform. The red line is the pitch it found.
What the machine kept
The notes as bars, the pitch bend through them in red, the loudness along the bottom. It plays through a simple voice.
The example: 7 notes: C♯3, D3, F3, A3, G3, F♯3, E3. The loudness moves across 13 dB, and the pitch strays up to 0.9 semitones from the notes. The Phrase as JSON.
The example is one take of me singing, as it came.
Try a phrase of your own
Hum it, sing it or play it, one line at a time, for up to 30 seconds.
It all happens in this browser tab. Nothing is uploaded.
The gap
I repeated a phrase played with a synth sound, which I'd like to get away from: sound experiments on the web using unattractive sounds. I sang five notes and it heard seven. I started a semitone under the D, and that became a C♯ of its own. The fall off the end of the G became an F♯. Three of the five were about a third of a semitone flat. The rounding hid that from the notes, and the bend kept it. Nothing in the Phrase says which of these I meant.
A Phrase holds three things: a list of notes, a loudness curve and a pitch curve. Whatever else was in the recording has nowhere to go.
Kept:
- Which notes, in what order, and roughly when.
- The shape of the loudness. A swell on one held note comes back as one note and a curve that rises and falls.
- What the pitch did around the notes. Notes get rounded, the bend doesn't, so a phrase sung 30 cents flat stays 30 cents flat. Vibrato and slides stay too, and the voice plays them back. So does a quick ornament: too short to be a note, but still in the curve.
Lost:
- Timbre. There is no field for it. A hum and a cello with the same line and the same shape make the same Phrase, and what you hear on the right is a stand-in voice that gets brighter as it gets louder.
- Words, consonants and breath. Anything without a pitch leaves only loudness.
- A change of bow, or a note re-sung with no break. The notes run together, and only a dip in the curve remembers it.
- More than one line at a time. A double stop on A3 and E4 comes back as a single A2, a note nobody played: the two strings together repeat every 110th of a second, and that's what the tracker follows.
These lists come from measuring synthesised test sounds, where the right answer is known. The example above is the first real voice through it.
How it hears
The pitch tracker is YIN, about 30 lines of JavaScript and no model to download. It looks for how far along the sound repeats itself. Spotify's Basic Pitch would hear chords, but it brings a neural network to load, and a phrase hummed into a phone is one line. It follows a line from a cello's open C up to a soprano's top C. Loudness is measured every 10 milliseconds.
Where this is heading
Days 3, 4 and 5 are steps towards an overnight composition factory: I record a phrase on my phone, and by morning it has become a set of sketches to listen to. This step could run as ordinary code, because the same recording always gives the same notes and curves. A program watching the folder would convert each new recording and save the result beside it. The one judgement worth keeping is a flag: a recording with no notes found comes to me rather than going on to the next step.