An experimental web application that transforms the semantic "latent space" of geographical landmarks into multi-voice orchestral compositions. By calculating the difference between two cultural or historical entities—as represented by the summary snippet from their associated Wikipedia articles—the app generates a unique auditory signature of their relationship.

The application is built with a modern, on-device AI stack to ensure privacy, low latency, and offline-capable processing.

The core of the app is the Sentence Transformer model (Xenova/all-MiniLM-L6-v2).

• On-Device Inference: The NLP model is downloaded to the browser's cache and runs locally using ONNX Runtime.
• Raw Vectorization: Unlike typical similarity-search apps, we disable normalization. This allows the Magnitude of the vector to represent the "Semantic Density" or "Energy" of the text. A short, simple description produces a quiet, minimal piece; a dense, complex history produces a high-energy composition.
• The "Journey" Vector: By subtracting the Origin vector from the Target vector, we calculate a Difference Vector. This is the musical representation of the semantic distance between two places.

The MusicEngine uses a curated selection of orchestral samples from the Philharmonia Sound Sample Library, chosen for their ability to handle rapid, data-driven triggering.

• Bass: Bassoon (A#1) — Provides a woody, foundational anchor.
• Tenor: French Horn (A3) — Fills the mid-range with warm, brassy textures.
• Alto: Spiccato Violin (A3) — Uses "bounced" bowing to handle 16th-note data streams with rhythmic clarity.
• Soprano: Flute (A4) — Provides agile, airy melodic top-lines.
• Percussion: A metrical section (Timpani/Toms, Woodblock, Triangle, Tam-tam) that anchors the data to a perceptible human pulse.

In Transformer models, information is distributed. We divide the 384 dimensions into four "Voices," giving each instrument a unique semantic window:

• Bass (0-95): Broad semantic clusters.
• Tenor (96-191): Cultural and functional features. Maybe. Who knows?
• Alto (192-287): Historical and relational data. Maybe. Who knows?
• Soprano (288-383): Specific stylistic and descriptive details. Maybe. Who knows?

The first few dimensions of the vector are used as "Global Variables" that set the musical rules for each specific landmark:

• Time Signature ($v[0]$): Determines if the piece "Waltzes" (3/4), remains steady (4/4), or feels tense and asymmetric (5/4).
• Articulation ($v[1]$): Determines if the performance is Legato (smooth and flowing) or Staccato (short and prickly).
• Texture ($v[2]$): Determines if the voices move together in chords (Homophonic) or chase each other independently (Counterpoint).

To ensure the data doesn't sound like "random noise," the engine employs Metrical Anchoring:

• The Bassoon and Toms are locked to the "Downbeat" (The 1).
• The French Horn and Woodblock provide a steady backbeat pulse.
• The Violin and Flute are the "Data Streamers," triggering notes only when semantic values cross a specific threshold.

By using this app, you are exploring the Latent Space of Human Knowledge.

• Semantic Energy: Notice how a long, detailed history of an ancient ruin results in a faster BPM and denser orchestration than a small local park.
• The "Journey" (Difference Vector): When you compare two landmarks, the music represents what makes them different. If you compare two similar cathedrals, the result will be quiet and minimalist. If you compare a skyscraper to a cemetery, the music will be harmonically and rhythmically complex.
• Character Shifts: Look for how the "feel" of the rhythm changes. A specific landmark might trigger a 5/4 time signature, creating an "unbalanced" feeling that is unique to its semantic embedding.

• Browser: Modern Chrome, Firefox, or Edge.
• Audio: Headphones are highly recommended to hear the spatial and frequency separation of the orchestral voices.
• Initial Load: Requires a one-time download of the NLP model (~23MB) and the orchestral sample library.

1. Install dependencies: npm install

2. Run the app: npm run dev

Samples provided by the Philharmonia Orchestra, London.