There is no published benchmark for how well music recognition works on social short-form content. Vendors publish accuracy numbers for clean studio recordings; nobody publishes numbers for a fifteen-second Instagram clip with someone talking over the drop, or a photo of a record sleeve that contains no audio at all. We run a music identification system on exactly that kind of content, so we measured it. This post is the benchmark: the numbers, the method, and the caveats — including the one result that cuts against our own pitch.
The Headline Result
On 38 real Instagram posts submitted by TrackRadar users, analysed once each through the full production pipeline in a controlled A/B experiment (August 2026):
- Audio recognition alone: 32 tracks identified.
- Audio plus text and image analysis: 71 tracks identified.
- Overlap between the two: 3 tracks (~4%).
Adding caption parsing, comment reading, on-screen text extraction, and image analysis on top of audio fingerprinting increased the identified-track count by 121.9% on this content. The two methods barely found the same tracks: of 74 distinct tracks, only 3 were caught by both the audio leg and the text leg. On short-form social content, audio recognition and text extraction are not redundant checks on each other — they see different tracks.
On 6 of the 38 posts (15.8%), audio recognition returned nothing while the full pipeline returned at least one track.
Cross-Validation on the Full Population
A 38-post experiment is small, so we validated it a second, independent way: source attribution across every completed Instagram analysis from real users in the production database — 219 analyses that returned at least one track (November 2025 through August 2026, excluding staff and demo content). By that method, audio recognition contributed 224 tracks and text, image, and metadata extraction contributed 280 — a lift of +125.0% over audio alone.
Two methods, one experimental and one observational, landing within three percentage points of each other. As a robustness check we also ran the same experiment on a separate corpus of 59 Instagram posts drawn from search-driven traffic to our free tools (a content sample, not a user-demand sample): the lift there was +40.4%, with the same near-zero overlap pattern. The size of the lift varies with the content mix; the direction does not.
How Often Everything Fails
The honest flip side: 37.0% of Instagram analyses return zero tracks even with every method running (351 real-user analyses). For comparison, the same system's zero-rate is 3.5% on SoundCloud (796 analyses) and 24.8% on YouTube (101 analyses). Instagram is the hardest surface we analyse, and a third of it is currently beyond every technique we have. Anyone claiming they can identify the music in every Instagram post is overclaiming — this number is what the honest ceiling looks like today.
The Result That Cuts the Other Way
This benchmark is not evidence that audio fingerprinting is weak. Aggregated across every platform we support, audio recognition produced 95.4% of all 12,387 tracks identified for real users. Long SoundCloud and Mixcloud mixes dominate track volume, they fingerprint well, and text extraction is a rounding error there. Audio fingerprinting is the workhorse of music identification.
The gap is specific to short-form social content, and the reasons are structural rather than fixable: the audio is compressed and often overlaid with speech; clips are shorter than the clean window recognition needs; and a large share of music posts are stills — a photo of a vinyl sleeve, a flyer, a tracklist screenshot — which contain no audio to recognise. On that content the track ID lives in the pixels and the caption, not the waveform.
Where the Text-Derived IDs Actually Come From
The 280 text-derived tracks in the real-user population break down by source as follows:
- Post captions — 114 tracks (40.7%)
- Track IDs visible on screen, read by a vision model — 105 tracks (37.5%)
- OCR on caption imagery — 28 tracks (10.0%)
- Instagram's own music metadata — 10 tracks (3.6%)
- Web detection on imagery — 10 tracks (3.6%)
- Vinyl labels read from photos — 8 tracks (2.9%)
- Video overlay OCR — 3 tracks (1.1%)
- Comments — 2 tracks (0.7%)
Two humble sources — the caption and text sitting visibly on screen — carry 78% of the text leg. Much of the value is not exotic machine learning; it is systematically reading what the poster already wrote.
Method
Design. Each post in the A/B experiment was scraped exactly once, and the pre-deduplication track candidates were captured. The production deduplication function was then run on two subsets: all candidates (full pipeline) and audio-derived candidates only. The difference is the number of tracks that exist solely because the text and image legs ran. One scrape avoids double cost, cache hits, and the fact that scraping and recognition services do not return identical results across repeated calls.
Why the lift is a floor. Deduplication resolves conflicts by confidence, and audio sources carry higher confidence than nearly all text sources. A track surviving with a text source therefore means the audio leg did not produce it at equal or higher confidence. Removing the two text sources that can outrank audio still leaves a lift of +74% — the conclusion does not depend on the tie-breaking.
Populations. Real-user figures cover completed analyses from registered users, excluding staff accounts and demo content. The tools corpus is search-driven traffic to free tool pages: valid as a technical sample of Instagram content, never citable as user demand. Every number above states which population it comes from.
Caveats. Recognition is stochastic and catalogues grow, so re-running the same URLs on a different day shifts individual results (the experiment's zero-rate, 15.8%, is lower than the historical population's 21.5% for exactly this reason — we cite the experiment's own numbers only for the experiment). Text-derived labels are text-asserted and platform-linked, not human-verified. Aggregate figures include every recognition provider we use; this benchmark compares modalities within one system, and is not a comparison between named recognition vendors or consumer apps.
For Researchers and Recognition Vendors
Behind this benchmark sits an evaluation dataset: real short-form posts where audio recognition returned nothing and a text or image method produced a platform-linked track ID, each case carrying its source and confidence. As far as we know, no public test set for music recognition on social short-form content exists. If you work on audio recognition, music information retrieval, or evaluation and want to talk about the data or the method, email info@trackradar.ai — happy to share detail beyond what fits in a blog post.
Try It on a Real Post
The system behind these numbers is live: paste an Instagram URL into the Instagram song finder and it runs the full pipeline from this benchmark — audio recognition plus caption, comment, and image analysis. For background reading, see how to identify tracks from Instagram and what AI track identification actually does.