TangoTube
From the desk

Clips are here

Create short practice clips from any video. Loop movements, tag techniques, and build your personal library.

Try Clips
Engineering notesVolume 2 of 8 · The music
Volume 2

Naming the Recording

A tango title names a song and a band. It almost never names the recording, and the same band often recorded the same song twice, decades apart. This is how TangoTube says exactly as much as it knows about the music in a video, and never a word more.

recordings on the shelf
16,629recordings on the shelf
of song + band pairs are one take
93.7%of song + band pairs are one take
need someone to listen
3.6%need someone to listen
recordings we can listen against
5,678recordings we can listen against
videos with a named recording
51,795videos with a named recording

Part 01

A song is not a recording

a video arrives“Sobre el pucho”by D’Arienzothe shelf answers19411953the judgewhich take?a wrong year is worsethan none at allPoint at the song and you lose the band. Point at one take and you invent a year.
The video speaksA song and a band. That is all a title ever gives.
The shelf answers with twoD'Arienzo recorded it in 1941 and 1953. Dancers can tell them apart.
The judge will not guessBoth obvious answers are wrong in a way that shows on the page.

Dancers do not choose songs; they choose recordings. Pugliese’s 1952 La Tupungatina and D’Arienzo’s 1948 are the same poem and completely different dances. So the question is not “what song is this?” but “what is the most specific true thing we can say about the music, and how do we stop the page saying more?”

Part 02

The shelf is not the notebook

TangoCloudthe librarianthe shelf16,629 real recordingswe only read itthe notebookwhat we believe about each videoand how sure we are“I believe: D’Arienzo, 1941”points at one itemthe matchermay never add a book to the shelfWhen nothing fits, the notebook says “nothing yet” and keeps the clue.
The shelfThe real list of recordings, kept outside, read-only.
The notebookOur beliefs about each video, each with a confidence and a reason.
The one ruleThe matcher can write in the notebook. It can never put a book on the shelf.

The first version had one list and let the matcher fill it. After three years there were 273,041 “songs” and only 3,185 had ever been played in a video. The rest were guesses dressed as facts. Separating the shelf from the notebook is what stopped that.

Part 03

Count before you build

13,216 song + band pairs on the shelfone recording · 93.7%Knowing the song and the band is knowing the take.Pin it straight to the recording.A named singer, or “only one is from 1935–1955”, settles most of the rest.Four danced recordings in five are from those years.474 pairs, 3.6%, need someone to listen. That is the whole hard problem.
GreenFor 12,385 pairs there is exactly one take. No ambiguity to model.
AmberA singer credit or the golden-age hint settles 357 more.
Grey474 pairs need ears. The listening system in scene 5 is sized for them.

We nearly built a third thing between the song and the recording to hold the ambiguous cases. One count showed it would serve 3.6% of the shelf. So there is no third table: a song-and-band key groups recordings, and a small pile waits for ears.

Part 04

Say exactly what you know

identifier or fingerprintthe take“Recorded 12 September 1941”song + band, two takes existsong + band“D’Arienzo recorded this in 1941 and 1953”a note records the guess and its alternativestitle only, no band namedthe songcomposer and lyricist shown · no band, no dateThe evidence decides the depth. The page reads only as deep as the entry allows.
The takeAn identifier, a fingerprint, or a pair that exists once on the shelf.
The pairTwo takes remain. The page names both years and never a day.
The songThe title named the poem and nothing else. No band, no date.

When we lean on the golden-age hint to show a likely take, we write down that we guessed: which hint, when, and the alternatives. Deleting that note is how a guess becomes a fact. Nothing gets flipped; something gets removed.

Part 05

Check the drawer, then listen to your own shelf

the drawerclues stored years agosame identifierthe shelffreetens of thousands of takes namedby a lookup between two listsonly what is left goes here60 seconds of soundour own shelf · 5,678 recordingsonly if very surea wrong take is worsethan no take
The drawer firstIdentifiers stored years ago match the shelf's own. A lookup, not a project.
Then the lensShazam-style listening, against our own 5,678 recordings, not a rented database.
Same judgeA match is one more clue. Applied only when very sure; never over a person's word.

Commercial listening services are weakest exactly where tango lives, in old transfers of old records. Our shelf covers 13 of the 15 most-danced recordings. And “music, but nothing on our shelf” is not a failure: it means a live orchestra, and it becomes a filter of its own.

Take-aways, without the tango

  1. 01Keep the authoritative list and your beliefs about it apart. Never let the matcher write the list.
  2. 02Count the ambiguity before you model it.
  3. 03Say exactly as much as the evidence supports, and let the interface read only that deep.
  4. 04Write guesses down as notes you can delete, not flags you must flip.
  5. 05Check what you already stored before buying signal. Listen last, and to your own shelf.

The Cabeceo

We've caught your eye across the milonga. Like any good tango, this starts with consent — nothing optional runs until you say so. What we store