How SOONOS Analyzes a Full DJ Mix

The set is analysed end to end

A DJ set is judged on how it unfolds. SOONOS therefore works on the full recording: the warm-up, the build, the peak and the way the set lands. Nothing is sampled at random and no score is extrapolated from a 30-second window. When part of a file cannot be measured with confidence, the report says how much of the set was covered instead of guessing.

Audio pre-analysis

The file is decoded and normalised before anything is scored. The engine checks the format, length and integrity of the recording, builds the waveform, and prepares the frames used by every later stage. Files that are too short, corrupted or unreadable never produce a score — they are marked incomplete.

Measured signals (DSP)

These values come from the audio itself. They are measurements, not opinions.

• Beat-grid behaviour: grid breaks over time and the median beat offset between the two records in play.
• Tempo: BPM detection and tempo stability across the set.
• Loudness and dynamics: level consistency, and limiter behaviour when a real ceiling is measured.
• Energy: an energy curve sampled across the whole set, used for structure and for the crowd model.

Interpretive analysis

A multi-agent listening pass reads musical intent, which no measurement can express: track selection and programming, storytelling across the set, energy management, surprise, and the identity that makes the set recognisable. This layer interprets — it does not measure — and the report labels it as such.

Transitions

Transitions are detected in the audio and analysed one by one. Each one carries a timestamp, a type or technique when it can be identified, a score, and a short comment on what happened. Repeated behaviour across the set — long static loops, recurring EQ clashes, the same exit every time — is reported as a pattern rather than as an isolated event.

Timestamps and audio evidence

Every insight in a public report is anchored to a moment in the set. Where the original recording is publicly hosted and supports it, the report links straight to that timestamp so the moment can be heard. Insights are tagged with how they were produced — measured (DSP) or listening pass — so the origin of each statement is visible.

How measured signals cap interpretive scores

The two layers are merged, and the measured layer wins on execution. If the audio shows beat-grid breaks, tempo drift or a hard limiter, that measurement sets a ceiling on the final score even when the listening layer is enthusiastic. This is why a global score is not the average of the visible dimensions: a single measured execution problem can hold the whole set down. When a ceiling is applied, the report names it.

Coverage and confidence

Coverage is the share of the set the engine could measure with full confidence. Encoding artefacts, heavy compression, very low-quality sources or unstable recordings reduce it. Low coverage does not invent a score — it lowers the confidence attached to the conclusions, and a report that cannot be measured well enough stays out of the public index.

What SOONOS can identify

• Beat-grid and tempo execution problems, with the moments where they happen.
• Loudness and dynamics behaviour across the set.
• Transition quality, technique and recurring habits.
• Energy structure: where the set builds, where it drops, where attention is likely lost.
• Programming, storytelling and identity signals from the listening pass.

What SOONOS cannot guarantee

• It does not know what the room actually did on the night — no real crowd data is used.
• It does not read intent: a deliberate rough cut and a mistake can measure the same.
• It does not judge taste. A style that is not your own is not a lower score by itself.
• It cannot compensate for a poor recording: a bad capture limits what can be measured.
• Scores vary slightly between runs on the same file; a few points of movement is normal.

Why the report is guidance, not truth

The measured layer is factual, the listening layer is an informed reading, and the set was played for a room SOONOS never heard. Used well, a report tells you what to work on next and lets you prove the change on the next upload. Used as a verdict, it says more than the data supports.

Real examples

Three public analyses produced with exactly the process described above:

Go further