{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreib7k4yp54ys73kk2whgzy4wltexumzdccspavqjbpwey5cin2wgou",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3movckznzmgm2"
  },
  "path": "/t/a-10kb-page-that-sings-text-can-you-decode-it-back-pitch-only-baseline-already-hits-43/176840#post_7",
  "publishedAt": "2026-06-22T15:14:28.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "loom carrier decoder",
    "loom booth v2 - carrier lab"
  ],
  "textContent": "Hadn’t checked in here in a bit since posting. Since my first post, the project has moved a lot, but in a way that mostly confirms your decomposition.\n\nTo answer the original question of, “can you decode it” the answer is yes, I think. You were spot on about needing more structured metadata. There’s now an updated booth, which adds “timestamp - alphabet code” pitch. This pitch iterates on the previous two pitches which were not differentiated enough to be fully recoverable. This one is, in theory. I created a decoder that can decipher raw mp3 recordings of the booth singing. Its sensitive to audio quality, and I haven’t gotten a perfect score yet. Been able to surface results like this\n\n\"opera oohs and aahs\n\nraw mp3 run through our decoder of oohs and aahs: brunsob� won tha�.�vb�e.jx��!db� finished the joa�; a�vidv?�ce: lac�.?,vjd� jumper,�b�i�db�.\n\nOG text: Brunson won the game Brunson finished the job; evidence: lap 13 elbow jumper - nominal.\"\n\n* * *\n\nloom carrier decoder\n\nAdditionally, I created loom booth v2 - carrier lab to test outputting the tape in different formats than audio, such as mosaic, scatterplot, loom pixel, varint byte stream, glyph stream, and ascii path tape.\n\nDecode the above if you want! There’s a field in the decoder for uploading a loom pixel, upload this png and decode it.\n\nSo the “audio → notes/motifs → text/alignment” framing aged well. The booth/carrier side is still basically a controlled synthetic decoding sandbox.\n\nCan only post two links per post so the reply to this thread will be the new singing dedicated booth: updated controls, same spirit. You can also lightly query individual questions against the deterministic retrieval engine. It will sing back its answers in whatever format you have configured, according to the loaded tape, and what it can instantly retrieve from Duckduckgo search.",
  "title": "A ~10KB page that sings text — can you decode it back? (pitch-only baseline already hits 43%)"
}