Upload README.md with huggingface_hub

README.md

@@ -8,30 +8,30 @@ tags:
 - structure
 ---
 
-# Structural Derivation: A Model for Analyzing Musical Coherence
+# Structural Derivation: A Model & Metric for Musical Structure Analysis
 
-This repository contains a model trained to understand the internal structure of music. It can be used to generate two novel metrics for any given audio file: the **Structural Derivation Consistency (SDC)** score and the **Structural Diversity** score.
+This repository contains a model for learning structural embeddings of music by contrasting segments from the same song against segments from different songs. It can be used to generate two novel metrics for any given audio file: the **Structural Derivation Consistency (SDC)** score and the **Structural Diversity** score.
 
-These metrics are particularly useful for quantifying the high-level coherence of a musical piece and can effectively distinguish between human-composed music and music generated by AI models that may exhibit "thematic drift."
+These metrics are particularly useful for quantifying the high-level thematic structural consistency and structural diversity of a musical piece. Together, these metrics provide a lens on both the cohesion and variation of a track’s structure, distinguishing human-composed music (typically higher consistency and balanced diversity) from AI-generated music (often lower consistency or uneven diversity due to thematic drifts).
 
 ---
 
 ## 🎵 Key Metrics
 
-The model analyzes a song by first splitting it into segments (e.g., 20 seconds each) and generating a high-level embedding for each segment. These embeddings are then used to compute the following scores.
+The model analyzes a song by first splitting it into segments (e.g., 10-20 seconds each) and generating a high-level embedding for each segment. These embeddings are then used to compute the following scores.
 
 ### 1. Structural Derivation Consistency (SDC) Score
 
-The SDC score quantifies a song's **structural and thematic integrity**. It serves as a proxy for **musical narrative consistency**, measuring how well a song "remembers" and develops the musical ideas presented in its opening segment.
+The SDC score quantifies a song's **structural and thematic integrity**. It serves as a proxy for **musical narrative consistency**, measuring how well a song "remembers" and develops the musical ideas presented in its opening segment. In other words, it reflects how structurally coherent and thematically consistent the song is in embedding space.
 
 * **How it works:** The score is the average cosine similarity between the embedding of the first segment (S1) and every subsequent segment (S2, S3, ..., SN).
 * **Interpretation:**
-    * A **high SDC score** suggests the piece is thematically coherent. Like a well-written story, it builds upon its initial premise. This is characteristic of human-composed music, which often uses repetition, variation, and development of a core motif.
+    * A **high SDC score** suggests the piece is thematically consistent. Like a well-written story, it builds upon its initial premise. This is characteristic of human-composed music, which often uses repetition, variation, and development of a core motif or theme.
     * A **low SDC score** indicates potential **"thematic drift."** The song's later sections diverge significantly from its beginning, lacking a clear connection to the initial musical statement. This can be a trait of AI-generated music that struggles with long-form compositional structure.
 
 ### 2. Structural Diversity Score
 
-The Structural Diversity score measures the **variety of distinct musical ideas** within a single piece of music. It complements the SDC score by providing insight into the song's complexity and repetitiveness.
+The Structural Diversity score measures the **variety of distinct musical ideas** within a single piece of music. It complements the SDC score by providing insight into the song's complexity and repetitiveness. In other words, it reflects the variety and richness of internal structures by capturing how different the segments are from one another.
 
 * **How it works:** The score is calculated using the [Vendi Score](https://github.com/verga11/vendi-score) on the similarity matrix derived from all segment embeddings.
 * **Interpretation:**