How the Songbrain Virality Score Actually Works
May 18, 2026 · 7 min read
Most "AI music analysis" tools give you a number and a vibe. You don't know what went into it, you don't know what would change it, and you definitely don't know whether it's telling you about your song or about itself.
We'd rather just show you the pipeline. The Virality Score is the output of 10 separate AI models running in sequence, fused at the end. Here's exactly what each one does and how their votes combine.
The pipeline, end to end
When you upload a song, it goes through this sequence — sequentially, because every stage builds on the previous one's output. Total time: about 60 seconds for a 3-minute track.
Tempo, key, loudness (LUFS), dynamic range, spectral centroid, frequency balance — the same vocabulary mastering engineers use.
Two independent neural genre detectors voting. MusicNN gives a confident top-1, CLAP cross-checks via natural-language embedding so we catch crossover cases (e.g. dark-pop that's really hyperpop).
Once genre is locked, a second pass routes to 1 of 37 curated subgenres — the granularity that actually maps to discovery (phonk ≠ trap, shoegaze ≠ indie-rock).
Full timestamped lyric output. Confidence-scored per segment so the downstream evaluator knows what to trust.
Emotional delivery, hook clarity, repetition pattern, originality. Compared against genre-matched chart references — country lyrics aren't graded against drill.
Identifies the production stack — what's actually in the mix, layered or sparse, live or sample-based. Feeds into the "does this sound like the current chart productions in your genre" check.
The 3–15 second windows with the strongest hook signal. Cross-validated by spectral energy peaks, vocal density, and lyric-payoff alignment.
We pull viral TikToks per subgenre weekly and index the exact second of the song people clip. Your best moments get matched against current viral windows — not against last quarter.
Combines all the above into one number per dimension (hook strength, production quality, trend fit, lyric impact, energy contour). This is the "fusion" step — no single model dominates.
Final 0–100 number. Weights are tuned per genre — what makes a 90 in metalcore is not what makes a 90 in K-pop. The score reflects viral potential within your genre's viral language, not a universal one.
What gets weighted heavily — and what doesn't
The fusion step doesn't treat all models equally. Some signals predict virality much better than others, and some are easy to game (which makes them worse signals, not better). Here's the rough weighting:
| Dimension | Weight | Why |
|---|---|---|
| Best Moments quality | Heavy | Strongest signal — hook strength + placement. |
| Trend alignment | Heavy | How well your sound maps to current viral windows in your subgenre. |
| Production quality | Medium | LUFS, dynamic range, mix balance vs current chart references. |
| Energy contour | Medium | Does the track build → drop → resolve in a way that holds attention? |
| Lyric impact | Medium | Hook clarity, originality, emotional payoff alignment with the music. |
| Catchiness fingerprint | Light | Repetition + melodic memorability — important but easily overfit if weighted too high. |
Why the score is genre-relative
A 90 in pop and a 90 in shoegaze do not mean the same thing in absolute production terms — they mean both tracks are equally well positioned to go viral inside their own genre's viral language. A shoegaze track that sounds like pop would be a failure as shoegaze; a pop track with shoegaze production would confuse the algorithm.
So the fusion model carries genre context. The same set of audio features produces different scores depending on what genre routing placed the track in. If the genre is misclassified upstream (rare, but it happens with hybrid tracks), the score is wrong — and you'll see this on your report because the genre label will look off.
Why the trend match matters more than people expect
Virality has a time dimension that most analyses ignore. A track that would have gone viral in February might not move in May because the algorithm trains on different windows.
That's what stage 08 (Apify weekly scrape) is for. We pull the top viral TikToks per subgenre every week and index the exact second of the song people clip — sound-IDs, ISRCs, lyric snippets, hashtags. Your Best Moments get compared to this week's viral windows, not to a frozen historical reference. If shoegaze-on-TikTok is currently obsessed with the 1:45–1:58 window pattern, that's what your shoegaze track is graded against this week.
What the score does not do
- It does not predict revenue. Viral ≠ monetizable.
- It does not account for promotion, posting cadence, audience size, label support. Pure song-level signal.
- It does not punish "non-trendy" songs — if your genre's viral window happens to favor longer hooks or slower builds, the score adapts.
- It is not a creative judgment. A 60 doesn't mean the song is bad. It means in the current viral language of its genre, it would have to fight harder for placement.
What you can do with it
The score is most useful when you compare it across versions of the same track. Upload v1, see what falls short. Adjust the hook placement, tighten the mix, restructure the intro. Re-upload as v2. Watch the score move. That's the feedback loop the product is built for — it's an iteration tool, not a verdict.
See your score in 60 seconds.
500 free credits on signup. Auto-OG Early Access during the launch window.
Analyze Your Song — Free →