Experience lab
Score Lens
See the structure inside the score.
The Music Analyzer deterministically gathers factual evidence and reproducible measurements from computer-readable orchestral scores. Score Lens tests a second layer: whether a controlled, human-calibrated dictionary can map combinations of those signals to useful natural-language descriptions of perception without presenting interpretation as fact. An LLM can then act as an interface to the validated data, while the listener inspects the evidence, hears the passage and makes the final judgement.
Conceived as part of the next markslater composer website experience; showcased here as work in progress.
The opportunity
Designed for Composers, orchestrators, music technologists and curious listeners
Computer-readable scores can reveal instruments, tempo, dynamics, activity and other factual conditions, but those facts do not automatically explain perception. Broad descriptive language has the opposite problem when the evidence behind it disappears. Score Lens keeps the two connected but visibly distinct.
Working now
- ✓Controlled natural-language search across 28 analysed passages from three selected tracks
- ✓Cue-accurate Tone.js audio playback with a persistent waveform player
- ✓Deterministic matching over prepared analysis and reviewed descriptive tags
- ✓MIDI-derived segmentation and musical measurements with MusicXML score context where available
Next on the roadmap
- 01Make source evidence, derived measurements, perceptual mappings and LLM narration visibly distinct in the interface
- 02Expand the controlled perceptual dictionary and test its mappings through human review across more cues
- 03Connect a conversational service that can query and explain validated results without inventing analytical evidence
- 04Bring the strongest interactions into the future composer website
Project in brief
“What makes an orchestral passage feel tense, suspended, inevitable or finally resolved?
The Music Analyzer first does the less ambiguous work: gathering source evidence and deriving reproducible measurements from real recordings aligned to MIDI and MusicXML. It examines instrumentation, tempo, density, notated dynamics, texture, dissonance, energy, tension and change over time, with the authority of each source kept explicit.
A second layer tests one possible dictionary between combinations of those signals and human perceptual language. Terms such as suspended, urgent or calm are navigational hypotheses to be calibrated through listening, not emotions claimed as objective properties of the score.
Score Lens makes the factual evidence visible beside each match and plays the passage from its exact point in the recording. An LLM can provide a natural-language interface to query and explain the validated data, but it does not choose the evidence or make the final musical judgement.
The constraint is the point: interpretation remains connected to inspectable facts, and the human listener remains free to explore, disagree and decide whether the description is musically persuasive.”