Song recognition from the audio, not from the caption

What song recognition is listening for

The useful picture is a fingerprint, not a transcript. A short stretch of the recording is turned into a compact signature and compared with signatures of recordings the catalog already knows. The signature is built from how the sound is shaped over a few seconds: the pattern of peaks a person would not sit and label as notes. You do not see that signature. You see a title and an artist when one recording lines up, or a miss when none does.

That is why the caption is irrelevant even when it is correct, and useless when it is wrong. Text on the screen never enters the comparison. A sticker that says the song is a faster path only because you can search the words yourself. Recognition does not “confirm the sticker.” It ignores the sticker. If the sticker and the preview disagree, believe the preview and your ears.

It is also why a description fails. “Female vocal, 2010s, the one from the perfume ad” is language. The catalog is not searched with language here. Lyrics you type are language too. They belong in a normal web search, which find a Reel song by lyrics treats as its own job. A melody you sing is a new performance. It can be a good query for a hum search. It is a bad query for a fingerprint of a specific master.

The provider that runs the comparison is AudD. We ask it for Apple Music, Spotify, and Deezer metadata when it has a hit. From that we show the title, the artist, cover art when a catalog provides it, a short preview when one of those services returns one, and store links or honest search links. We do not add a confidence percentage, and we do not fill a miss with a nearest neighbor. One alignment, or no match.

“A few seconds of the record” is the practical bar the product is built around. A sting that is barely there often will not line up. A long video is not scanned from start to finish for every cue. A large upload is reduced in the browser to the first 30 seconds of audio before it is sent. You get one match for that stretch. If the song starts later, the signature is of the wrong stretch, which is a miss even when the video, taken as a whole, clearly contains the record. Cut the moment first. That cutting job is find a song in a video.

Why noise, talking, or a speed change breaks it

A fingerprint is brittle in a specific way. Small changes can still line up. Large changes do not. The catalog holds the released recording, not every phone video of it. That brittleness is ordinary song recognition, not a sign that the upload button is broken.

Noise and room sound add energy the master does not have. A voice memo pointed at a laptop speaker picks up the table, the fan, and anyone talking. The peaks move. Mild room tone on a clear chorus can still hit. A clip shot in a club, or a voice memo of a shop speaker with chatter on top, often will not. If a public link exists, the file is the better sample because it skips the room. It does not skip a noisy mix the creator already baked in.

A voiceover is the same problem with a face. Speech is loud, broadband, and right on top of the hook. The signature starts to describe the voice. Pasting the link still helps, because you avoid adding a second noisy room on top of that mix. It does not help when the creator ducked the music until almost nothing remains between sentences. Look for another public clip that plays a longer stretch of the same audio. The case-by-case version is song under a voiceover on a Reel.

Speed and pitch move the recording away from the master on purpose. A small speed-up, the kind that makes a chorus feel urgent, often still hits the original. You then see the original’s title, which is a successful recognition of the underlying record, not a separate “nightcore” row. A large pitch drop, a heavy time-stretch, or a remix that keeps four words and replaces the track usually misses. A mashup asks one signature to be two records. It comes back as one of them, or as neither. We will not split it by listening the way a person would. The edit-specific page is sped-up or remix songs.

Unreleased audio fails for a duller reason. There is no master in the catalog to align with. A private loop, a snippet that is not out yet, and a stem from a project file are not “harder fingerprints.” They are absent. Recognition correctly returns no match. Guessing a famous song that feels similar would be a lie. That limit is new or unreleased songs.

Online song recognition and a phone app

“Online” here means the browser form. You paste a link or upload a file. Nothing is installed. The phone can stay muted, because a public Instagram, TikTok, or X link is fetched as a page, not heard through the speaker. A YouTube or Shorts URL is not fetched. Upload a recording of that video, or paste a Reel or TikTok of the same clip. Which hosts are real is the song identifier. This section is only about why the channel matters.

	This site, in the browser	A mic app (Shazam and the like)	A hum search (Google hum-to-search)
Input	Public Reel, TikTok, or X link, or an uploaded audio or video file	Whatever the microphone hears while the song plays	A melody you hum, sing, or whistle
Muted autoplay	Fine, if the public link opens or you have the file	Fails. Silence is the sample	Irrelevant. You supply the melody from memory
Best moment to use it	You can copy a link, or the file is already on the device	A speaker in the room, and no link you are allowed to copy	The clip is gone and the tune is still in your head
What it cannot repair	A mix that buried the record, a private URL, a hum	A muted phone, a two-second hook, a noisy room	It is not identifying the master you heard. It is searching a melody
Install	None	The app, plus microphone permission	The Google app or Google Search

A browser extension that listens to a tab, such as aha-music, is a third shape. It hears audio that is actually playing in the browser. A muted Reel gives it nothing. We do not run that extension, and we do not pass it your URL. If the tab is already making sound, their product can listen. If you have the link and the phone is muted, paste the link here.

None of this is a claim that one column “wins.” There is no accuracy figure on this page because we do not publish one, and we will not invent one for the other products. Pick the column whose input you actually have. The mic path for a Reel, when you truly cannot copy a link, is Shazam alternatives for Reels.

What leaves your browser

A link check sends the URL so we can see whether the public page opens. That check does not name the song, and it does not require an account. Recognition, after you pay, sends audio to AudD: either the media fetched from a public link, or the file you uploaded. A file over about 9 MB is turned, on your device, into a mono 16 kHz WAV of the first 30 seconds before that upload. The video image is not what the match needs. Above about 600 MB, or if the browser cannot decode the file, that shrink does not happen and an over-limit file is refused.

We cache a recognition result so the same bytes are not treated as a brand-new question every time. The cache key for an upload is a hash of the file. The stored value is the result, a title and artist or the error, not another copy of the audio filed under your name. That is not a promise about how long AudD keeps the audio they received to build the match. Their copy is theirs. Our privacy policy is the list of what we process: the URL, a visitor id, a hashed IP used for abuse limits, the checkout email and entitlement cookies if you pay, and product analytics that do not include the Reel URL, the file name, or that email. Read privacy for the actual list. We do not sell personal data. We do not open private Instagram accounts.

Paying is handled by Waffo, the merchant of record. We do not see the full card number. Access follows the checkout email, which you can restore on another browser at Restore. None of that changes what the fingerprint does. It only decides whether the comparison is allowed to run.

Do this

1. If the title is already printed, stop. A sticker, a sound page, or a filename that is the song does not need a fingerprint. Recognition is for the case where the audio is the only reliable source.
2. Give it the recording, not the room and not a performance. Paste a public Reel, TikTok, or X link. Upload an MP3, M4A, WAV, OGG, or a video file when that is what you have. Do not hum. Do not paste a Shorts URL. The upload path for a file you already saved is find the song from an audio file.
3. Put the song in the stretch that will be heard. On a large file that means the first 30 seconds. On a long video, cut 15 to 30 seconds around the moment instead of uploading the whole thing and hoping. One match returns.
4. Pay, then use the preview as the check. The match runs after you pay. A no match uses one credit and means the signature did not line up. A failed fetch is not charged. Play the preview when one is offered before you save anything. How to decide the preview is wrong is is the Reel song match right.

A wrong match and an empty one

An empty result is not a bug by default. Empty song recognition means no catalog signature lined up with the sample. Change something that changes the sample: a cleaner clip, a longer stretch of the same audio, the unedited hook instead of the nightcore. Repeating the identical file spends another credit on the same answer. The retry list is when a Reel song is not found.

A wrong title is a real alignment with a different recording. Samples, covers, and mashups do this. The preview is the test. If the preview is not what you heard, do not treat the title as close enough. We do not show a second guess. Background music and original-audio stickers, which are about the kind of audio rather than the mechanism, are the music identifier.

A rejection is a third outcome, and it is not a miss. The wrong host, a Story, a private post, or a share link that never resolves fails before AudD names anything. You are not charged for that. Fix the input. Do not read it as “the song is unknown.”

What this page is not

It is not the source list. Which links paste and which files upload is the song identifier. It is not the background-music and ISRC walkthrough. That is the music identifier, including the free ISRC lookup you run only after a title exists. This page is the mechanism: what a fingerprint can hear, why the usual edits break it, how that differs from a mic app or a hum, and what leaves the browser when you run it.

The form is the same form as the homepage. It does not recognize a song before you pay, and this page does not change the plans. Amounts are on pricing.

Questions