Methodology

Every mix runs through the same deterministic pipeline before a single score is generated. Nothing is eyeballed — the numbers come first, the narration comes after.

Audio file
  → Decoding (Web Audio API / FFmpeg)
  → Low-level DSP features (FFT, RMS, autocorrelation, chroma)
  → Event detection (transitions, drops, breaks, phrases)
  → Multi-pillar scoring (6 dimensions, 0–100)
  → Calibration & Bayesian fusion (technique 70% + perception 30%)
  → DJ DNA fingerprint matching (8-D cosine similarity)
  → LLM narration (coaching, NOT scoring)
  → Persisted report (DB is the source of truth)

• FFT (Fast Fourier Transform) — breaks the audio into its frequency components to track spectral energy over time.
• RMS (Root Mean Square) — measures loudness and energy envelope across the track.
• Autocorrelation — detects tempo and rhythmic periodicity, the backbone of BPM tracking.
• Chroma features — extract harmonic and key content to inform harmonic mixing analysis.
• Spectral flux — flags sudden changes in frequency content, useful for catching transitions.

Raw signal features are fed into detectors that identify the structural moments of a set — where it builds, where it drops, and where the DJ moves between tracks.

• Transitions — identified by overlapping energy and frequency shifts between two tracks.
• Phrase alignment — checks whether transitions land on musical phrase boundaries using the phrase_alignment model, rewarding mixes that respect the 8/16/32-bar structure.
• Drops — detected as sharp energy surges following a build-up section.
• Breaks — flagged where energy and rhythmic density drop off significantly.

Every mix is scored across six independent pillars. Weighting adapts based on set type and set length so a warm-up set isn't judged by peak-time standards.

1. Technical Mixing — beatmatching accuracy, EQ handling, and transition cleanliness.
2. Harmonic Mixing — key compatibility between consecutive tracks.
3. Track Selection — variety, coherence, and quality of the tracklist itself.
4. Energy Management — how well the set builds, sustains, and releases energy over time.
5. Creativity — use of effects, layering, and unexpected but effective moves.
6. Structure & Flow — overall pacing and arc of the set from open to close.

• Technique score (70%) — the objective, measurable half of the equation.
• Perception score (30%) — models how a crowd or listener would experience the set.
• Calibration — scores are normalized against thousands of reference mixes so 80 always means 80.
• Bayesian fusion — combines both signals into a single, weighted overall score.

Beyond the score, we extract an 8-dimensional fingerprint that captures your style as a DJ — not just how well you mixed, but how you mix.

1. Tempo range and drift
2. Transition style and length
3. Genre and key diversity
4. Energy curve shape
5. Effects usage frequency
6. Track selection patterns
7. Phrase alignment consistency
8. Set pacing signature

These 8 dimensions are compared using cosine similarity, letting us match your style to reference DJs and track how your DNA evolves over time.

LLM Narration

The language model never touches the numbers. Its only job is to explain why. the scores landed where they did, in plain language a DJ actually wants to read.

• It receives the final scores and detected events as structured input.
• It writes coaching feedback grounded strictly in that data.
• It cannot inflate, deflate, or override any score.

This separation means the overall_score you see is always deterministic and reproducible, even if the wording around it varies.