Dataset Viewer
Auto-converted to Parquet Duplicate
audio
audioduration (s)
0.62
5.27
text
stringlengths
7
34
motif
listlengths
2
6
you brake crew fast
[ [ 1 ], [ 6, 6, 3 ], [ 5, 1 ], [ 5, 5 ] ]
right beats
[ [ 3, 0, 4 ], [ 4, 0, 3 ] ]
now fast close win
[ [ 4 ], [ 5, 5 ], [ 2, 6, 3 ], [ 6 ] ]
heart lead
[ [ 1, 0, 1 ], [ 6, 2 ] ]
car three heart up lead
[ [ 1 ], [ 2, 2, 0 ], [ 1, 0, 1 ], [ 4, 0, 4 ], [ 6, 2 ] ]
down high
[ [ 2, 3 ], [ 5, 5 ] ]
rival rival now
[ [ 4, 0, 3 ], [ 4, 0, 3 ], [ 4 ] ]
down now
[ [ 2, 3 ], [ 4 ] ]
fast lose red up low
[ [ 5, 5 ], [ 1, 6 ], [ 3 ], [ 4, 0, 4 ], [ 3 ] ]
speed you right high
[ [ 1, 1, 2 ], [ 1 ], [ 3, 0, 4 ], [ 5, 5 ] ]
night up fastest stop beats now
[ [ 0, 4, 5 ], [ 4, 0, 4 ], [ 1, 1 ], [ 3, 5 ], [ 4, 0, 3 ], [ 4 ] ]
rival go brah beats up sharp
[ [ 4, 0, 3 ], [ 2, 4, 6 ], [ 6, 0 ], [ 4, 0, 3 ], [ 4, 0, 4 ], [ 4, 1, 5 ] ]
down fast
[ [ 2, 3 ], [ 5, 5 ] ]
win open fastest right lead city
[ [ 6 ], [ 5, 3 ], [ 1, 1 ], [ 3, 0, 4 ], [ 6, 2 ], [ 0, 4 ] ]
now rise brah night three
[ [ 4 ], [ 1, 3 ], [ 6, 0 ], [ 0, 4, 5 ], [ 2, 2, 0 ] ]
clean three heart
[ [ 1, 4, 4 ], [ 2, 2, 0 ], [ 1, 0, 1 ] ]
night left open loom gap fall
[ [ 0, 4, 5 ], [ 4, 3 ], [ 5, 3 ], [ 1, 6 ], [ 2 ], [ 6, 3 ] ]
then car high close
[ [ 3, 0 ], [ 1 ], [ 5, 5 ], [ 2, 6, 3 ] ]
move throttle loom you open
[ [ 5, 5 ], [ 6, 3, 3 ], [ 1, 6 ], [ 1 ], [ 5, 3 ] ]
red pit car throttle up
[ [ 3 ], [ 2 ], [ 1 ], [ 6, 3, 3 ], [ 4, 0, 4 ] ]
city loom clean echo then open
[ [ 0, 4 ], [ 1, 6 ], [ 1, 4, 4 ], [ 1, 5 ], [ 3, 0 ], [ 5, 3 ] ]
rise car heart lap hold clean
[ [ 1, 3 ], [ 1 ], [ 1, 0, 1 ], [ 5 ], [ 2, 4 ], [ 1, 4, 4 ] ]
fast gap
[ [ 5, 5 ], [ 2 ] ]
crew down fastest fall
[ [ 5, 1 ], [ 2, 3 ], [ 1, 1 ], [ 6, 3 ] ]
sharp turn pit echo
[ [ 4, 1, 5 ], [ 4, 4 ], [ 2 ], [ 1, 5 ] ]
rise echo
[ [ 1, 3 ], [ 1, 5 ] ]
fan high open
[ [ 5 ], [ 5, 5 ], [ 5, 3 ] ]
win race
[ [ 6 ], [ 2, 0 ] ]
apex three brake
[ [ 6, 2 ], [ 2, 2, 0 ], [ 6, 6, 3 ] ]
open heart move fall brake
[ [ 5, 3 ], [ 1, 0, 1 ], [ 5, 5 ], [ 6, 3 ], [ 6, 6, 3 ] ]
lap low lead up lap sharp
[ [ 5 ], [ 3 ], [ 6, 2 ], [ 4, 0, 4 ], [ 5 ], [ 4, 1, 5 ] ]
echo fastest turn lose you
[ [ 1, 5 ], [ 1, 1 ], [ 4, 4 ], [ 1, 6 ], [ 1 ] ]
stop you
[ [ 3, 5 ], [ 1 ] ]
pit lose the
[ [ 2 ], [ 1, 6 ], [ 3 ] ]
now stop dunk race the
[ [ 4 ], [ 3, 5 ], [ 5, 2 ], [ 2, 0 ], [ 3 ] ]
speed right brah
[ [ 1, 1, 2 ], [ 3, 0, 4 ], [ 6, 0 ] ]
down city low clean close fan
[ [ 2, 3 ], [ 0, 4 ], [ 3 ], [ 1, 4, 4 ], [ 2, 6, 3 ], [ 5 ] ]
rise fastest
[ [ 1, 3 ], [ 1, 1 ] ]
brake brake brake brake beats hold
[ [ 6, 6, 3 ], [ 6, 6, 3 ], [ 6, 6, 3 ], [ 6, 6, 3 ], [ 4, 0, 3 ], [ 2, 4 ] ]
fast go car win fall
[ [ 5, 5 ], [ 2, 4, 6 ], [ 1 ], [ 6 ], [ 6, 3 ] ]
high loom then
[ [ 5, 5 ], [ 1, 6 ], [ 3, 0 ] ]
beats the
[ [ 4, 0, 3 ], [ 3 ] ]
you right beats brah fan a
[ [ 1 ], [ 3, 0, 4 ], [ 4, 0, 3 ], [ 6, 0 ], [ 5 ], [ 5, 5 ] ]
win fan
[ [ 6 ], [ 5 ] ]
you rival dunk echo then
[ [ 1 ], [ 4, 0, 3 ], [ 5, 2 ], [ 1, 5 ], [ 3, 0 ] ]
hold high high open
[ [ 2, 4 ], [ 5, 5 ], [ 5, 5 ], [ 5, 3 ] ]
hold hold night heart you
[ [ 2, 4 ], [ 2, 4 ], [ 0, 4, 5 ], [ 1, 0, 1 ], [ 1 ] ]
apex loom
[ [ 6, 2 ], [ 1, 6 ] ]
hold clean move left
[ [ 2, 4 ], [ 1, 4, 4 ], [ 5, 5 ], [ 4, 3 ] ]
win left
[ [ 6 ], [ 4, 3 ] ]
you clean right a
[ [ 1 ], [ 1, 4, 4 ], [ 3, 0, 4 ], [ 5, 5 ] ]
night crew heart clean dunk left
[ [ 0, 4, 5 ], [ 5, 1 ], [ 1, 0, 1 ], [ 1, 4, 4 ], [ 5, 2 ], [ 4, 3 ] ]
move echo lose right
[ [ 5, 5 ], [ 1, 5 ], [ 1, 6 ], [ 3, 0, 4 ] ]
close loom rival lose fan throttle
[ [ 2, 6, 3 ], [ 1, 6 ], [ 4, 0, 3 ], [ 1, 6 ], [ 5 ], [ 6, 3, 3 ] ]
three brake apex
[ [ 2, 2, 0 ], [ 6, 6, 3 ], [ 6, 2 ] ]
go left open
[ [ 2, 4, 6 ], [ 4, 3 ], [ 5, 3 ] ]
gap a a lap
[ [ 2 ], [ 5, 5 ], [ 5, 5 ], [ 5 ] ]
dunk go clean then echo
[ [ 5, 2 ], [ 2, 4, 6 ], [ 1, 4, 4 ], [ 3, 0 ], [ 1, 5 ] ]
gap echo brah heart lose
[ [ 2 ], [ 1, 5 ], [ 6, 0 ], [ 1, 0, 1 ], [ 1, 6 ] ]
lose hold
[ [ 1, 6 ], [ 2, 4 ] ]
loom win hold
[ [ 1, 6 ], [ 6 ], [ 2, 4 ] ]
fan the hold crew echo crew
[ [ 5 ], [ 3 ], [ 2, 4 ], [ 5, 1 ], [ 1, 5 ], [ 5, 1 ] ]
pit high
[ [ 2 ], [ 5, 5 ] ]
sharp throttle go hold stop
[ [ 4, 1, 5 ], [ 6, 3, 3 ], [ 2, 4, 6 ], [ 2, 4 ], [ 3, 5 ] ]

loomspeech — a decode benchmark

Can a model listen to deterministically-encoded audio and recover the text it came from?

loom is a ~10KB, model-free codec: it turns text into sung audio by mapping each word to a fixed musical motif (1-3 notes on a pentatonic scale) sung on the word's own vowels. Same word -> same sound, forever, on any machine. So (text -> audio) labels are free and unlimited — this repo is a generator-backed benchmark, not a static dump.

The task

Given a .wav, predict the source text. Metrics: exact-line accuracy and position-wise word accuracy (score.py).

What's here

  • data/ — a playable sample (metadata.jsonl + wavs: file_name, text, motif).
  • encode.py — the encoder. Mint more: python encode.py --builtin --n 5000 --out data.
  • baseline.py — a non-ML floor: pitch-only, vocabulary-aware (~43% word accuracy). Beat it.
  • score.py — the scorer.
from datasets import load_dataset
ds = load_dataset("asleepyhimiko/loomspeech")   # audio + text + motif

Baseline & headroom

Pitch-only + known lexicon ~= 43% word / 20% exact-line. It throws the vowels away (which carry out-of-vocabulary words), assumes the lexicon, and dies under noise. A real model should beat all three.

The question

Near-100% decode proves a lossless, model-free code — meaning survives sound and returns. Whether a recoverable code is a language (human-learnable, conventional, productive) is the open question, not the claim. This benchmark measures the necessary condition.

Downloads last month
142