The numbers

Every claim on this site that can be measured links here: the number, how we measured it, the bar we set, and where it misses.

Search in words

6.5 of the top 10

results are the kind of sound you asked for, across 52 searches, with the single notes renamed so their names don’t give the answer away.

The claim: Describe it in words, and Mixel searches every sound in your folders, privately, on your Mac.
The bar: 8.5 of the top 10, set on the test Library’s own names, where Mixel reaches 8.2. The 6.5 above is the harder test, with the notes renamed. Not met yet.
How we measured: A test Library of 1,421 openly licensed sounds: 480 single notes from NSynth (CC BY 4.0), 521 drum one-shots from Freesound (CC0 and CC BY) and 420 loops from FSL10K (CC0 and CC BY). The notes were renamed so a name no longer says what instrument it is; the drum and loop names stay as their uploaders wrote them. Mixel 0.1.15’s search engine answers each search; a result counts when its label is the thing asked for. Precision at ten 0.65; 95% interval 0.54 to 0.75.
Where it misses: Loops are the hardest: 4.4 of the top 10 with the test Library’s own names.
A word nobody wrote down can’t be proven. With every file renamed, nothing is proven, every search ends in “No close match”, and the closest by sound are right 6.3 times in 10.

Your names count

60% of 14,337 sounds

in one producer’s Library (our founder’s) carry a word Mixel knows in their file or folder name. A file name you remember comes first in 96 searches of 100.

The claim: A sound’s file and folder names count as proof: “named kick”.
How we measured: Counted on that Library’s Mac with a script that reads a private copy of the Library and reports counts only: no names or paths left the Mac. 42 of 46 matches were checked by ear and were right. The remembered-name figure is from the test Library described under Search in words.
Where it misses: The other 40% have no such word. For those, only how they sound can order them, and Mixel marks them “sounds like”.
One Library, one producer’s naming habits.

Leave things out

1.9% slip through

of the results are what you said to leave out, across 16 “A, not B” searches. Without “not”, 10% would be.

The claim: Add “not vocal”, and sounds tagged, named or detected as vocal are left out, with a count.
The bar: 3% or less. Met.
How we measured: A test Library of 1,421 openly licensed sounds: 480 single notes from NSynth (CC BY 4.0), 521 drum one-shots from Freesound (CC0 and CC BY) and 420 loops from FSL10K (CC0 and CC BY). Each search names a thing and a thing to leave out; a result that is the left-out thing counts against it.
Where it misses: “Not” can only leave out what’s tagged, named or detected. With the notes’ names hidden, 9.4% slip through.

When it can’t prove a match

15 of 24

things missing from the test Library were answered with nothing shown as proven, as they should be. “No close match” itself showed for 8 of them.

The claim: When nothing in your Library is tagged, named or detected as what you asked for, Mixel says “No close match”, and still offers the closest by sound.
The bar: 22 of 24. Not met yet.
How we measured: A test Library of 1,421 openly licensed sounds: 480 single notes from NSynth (CC BY 4.0), 521 drum one-shots from Freesound (CC0 and CC BY) and 420 loops from FSL10K (CC0 and CC BY). 24 things the Library doesn’t hold were searched for; an answer passes when it shows 5 or fewer results as proven.
Where it misses: The misses are other sounds named with the word: a loop called “kick” or “bass” after the kick or the bass was taken out of it.

Find similar

7.5 of the 10 nearest

share the kind of sound you started from, for single notes and drum one-shots. For loops, 4.5.

The claim: Pick any sound and hear the ones nearest to it, across every folder.
How we measured: A test Library of 1,421 openly licensed sounds: 480 single notes from NSynth (CC BY 4.0), 521 drum one-shots from Freesound (CC0 and CC BY) and 420 loops from FSL10K (CC0 and CC BY). A neighbour counts when it shares the starting sound’s label. Run with our eval suite on the exact model files Mixel 0.1.15 ships (LAION CLAP, larger_clap_music_and_speech, Apache-2.0).
Where it misses: Leave out neighbours from the same pack, instrument or uploader, and the prototype with the same model gave 6.8 for notes, 3.8 for drum one-shots and 4.0 for loops.

Listening, on your Mac

2,662 sounds a minute

listened to on one Mac’s Neural Engine, at 0.38 joules a sound.

The claim: Mixel listens to your sounds with a model that runs on your Mac.
How we measured: Measured on our founder’s Mac, which has an M4 chip, with Mixel 0.1.14 listening on the Neural Engine alone.
Where it misses: One Mac. Older Macs will be slower, and we haven’t measured them yet.
The first time, Mixel downloads its sound model (416 MB) and prepares it for your Mac before it starts.

Numbers change as Mixel does; each says how and where it was measured. Found a number that doesn’t hold on your Library? Write to hello@0x1audio.com.