SongFinder identifies a song from a public link by listening to the audio, not by reading the label under the post. It downloads the media, cuts 15-second windows, and matches them against a music database. Everything below was checked against the running server on 2026-10-30, using the site's own host and direct HTTPS requests — no browser was involved, so anything that depends on a phone or a specific browser is explicitly marked as untested.
What happens to your link, in three steps
UNSUPPORTED_PLATFORM.That third step is why the tool behaves like Shazam and not like a database lookup: two uploads of the same song at different volumes should land on the same fingerprint, while a label lookup would depend on whatever text the creator typed.
Eight inputs, eight outcomes (2026-10-30)
| Input | Result | Time | What it shows |
|---|---|---|---|
Direct MP3, 8,945,229 B, audio/mpeg | result — a track name came back | 2.17 s | Direct file links do work. On 30 October the response reported sampledAtSeconds: 15, the single window that build used; since 4 November the search starts at the end of the file — see the window section below. |
| Direct MP3, 3 s long, behind a 301 redirect | no-match | 0.59 s | Redirects are followed. A three-second file reaches matching but carries too little to fingerprint. |
| Direct MP4, 5 s long, behind a 301 redirect | no-match | 1.29 s | Video files take the same route: the audio track is extracted, the video is discarded. |
An ordinary web page (example.com) | UNSUPPORTED_PLATFORM | 0.05 s | A page is not a file. Recognised by content type, refused in milliseconds. |
A .mp3 address that returns 404 | UNSUPPORTED_PLATFORM | 0.17 s | The extension lies: a dead link usually answers with an HTML error page, which is treated as a page, not a file. |
| Public Instagram Reel | result — a track name came back | 4.09 s | The platform route works and reports sampledAtSeconds: 0, i.e. from the start rather than second 15. |
| Public TikTok post | no-match | 3.92 s | Reaching matching is not the same as finding a song. A post with ambient or original audio can legitimately return nothing. |
| The same address, twenty-first request within the hour | HTTP 429, RATE_LIMITED | <0.1 s | Header Retry-After: 3600. Requests one to twenty each returned 200 — failed attempts count too. |
Read the table as one sample set, not as a benchmark. It is a small sample from one host, and the rate-limit row was re-measured on 5 November 2026; it says nothing about speed for visitors on other networks.
The six outcomes, and what to do about each
| Code | Meaning | Next move |
|---|---|---|
result | A snippet was matched. | Check the returned title against a 5-second listen. Fingerprint matches can land on a neighbouring release. |
no-match | Audio was read and sent; nothing was close enough. | Set a start time where the music is clean, or use a link that actually contains the track you hear in your head. |
UNSUPPORTED_PLATFORM | The address is not an accepted host, or it answers with a web page. | Use the post's own public URL. If a direct file link gives this, the link is probably dead. |
FETCH_FAILED | The server could not get the media: DNS failure, a redirect chain that broke, a 60-second timeout, or a response that is neither audio, video nor a page. | Open the link in a private window. If it downloads there, retry; if it shows a page, it is not a file link. |
NO_AUDIO | A file arrived but carried no usable audio track, or the extracted snippet was under 2 KB. | Silent video, a still image wrapped as MP4, or a corrupted file. Try a different source. |
OVER_LIMIT | Above 100 MiB by declared length, or above 10 minutes once probed. | Trim the section you need first. The limits apply to the source file, not to the snippet. |
Where the 15 seconds come from
Without a start time, more than one window is tried. Since 4 November 2026 the server cuts up to four windows and works backwards from the end of the file: the last 15 seconds first, then 75, 50 and 25 percent of the usable length (four windows for media 40 seconds or longer, three below that), stopping at the first match. The answer reports the window that matched as sampledAtSeconds. Re-measured on 6 November 2026: a 372.7-second MP3 matched at second 357, which is the last 15 seconds of that file, and the same file with a start time of 30 seconds matched at second 30. A start time still pins one window; it is clamped to the file length minus the snippet, so a value past the end slides back instead of failing.
The snippet itself is fixed: 15 seconds, audio only, mono, 96 kbps MP3 — the encoder flags come straight from the service's own extraction step (-ss, -t 15, -vn, -ac 1, -b:a 96k), so this is a description of the build, not an estimate. That is also why a "better" file does not buy a better match — a clean 15-second window beats a lossless file whose best 15 seconds are buried at minute nine.
When the answer is wrong, not missing
One of the samples above was a synthetic test track — a generated piece of music that is not a commercial release. It came back as a named pop record. The matcher returned its nearest neighbour rather than admitting there was nothing to return, which is normal behaviour for this class of tool and not a SongFinder quirk.
So treat a result as a lead, not a verdict: play the preview link that comes with it. If the preview does not match what you heard, the fingerprint landed on the wrong side of a similar-sounding track, and searching the artist plus a lyric fragment is usually faster than re-running the same link.
The limits that bite first
- Twenty attempts per address per hour. A failed attempt spends one. The twenty-first request in the hour gets
429withRetry-After: 3600— confirmed above, not estimated. - One shared daily pool for the whole site. It is not a thousand per person; when it is gone, everyone waits for the reset.
- 10 minutes / 100 MiB on a direct file, judged on the source file.
- No YouTube, no uploads, no humming. Those requests never reach the audio stage, so retrying them is wasted time. The full limits page lists all of them with the numbers.
Questions people actually ask
Does SongFinder read the music label on the post?
No. It downloads the audio and matches a fingerprint. A post labelled with a track can still come back empty, and a post with no label at all can still be identified.
Do I need to play the video?
No. The server fetches the media itself. Nothing plays on your device and nothing is uploaded from it.
Why does a link that worked yesterday fail today?
Either the post was deleted or made private, or you have used all twenty attempts for your address this hour. Both look similar on screen; the HTTP status tells them apart (HTTP 200 with a code versus HTTP 429).
Which links are refused before any audio is fetched?
YouTube links, Douyin links, a file on your phone, a hummed melody and any request to download the song. Those are different tools, not different settings.