| # Primitive Extractor | |
| ## Overview | |
| The Primitive Extractor is the first trainable module in the WindOne Sound V3 architecture. Its primary function is to compress and refine the high-dimensional audio features extracted by the frozen BEATs model into a lower-dimensional, highly structured "primitive" representation. | |
| ## Architecture & Flow | |
| - **Input Shape:** `[Batch, Time, 768]` (Raw features from BEATs) | |
| - **Output Shape:** `[Batch, Time, 512]` (Refined primitives) | |
| - **Total Parameters:** ~35 Million | |
| ### 1. Convolutional Stack | |
| The input is transposed and passed through three 1D Convolutional layers with decreasing kernel sizes (7, 5, and 3). Each convolution is followed by Group Normalization (32 groups) and a GELU activation function. This step captures local, short-range temporal patterns and acoustic textures in the audio. | |
| ### 2. Linear Projection | |
| The output of the conv stack (1728 dimensions) is projected down to 512 dimensions using a Linear layer, followed by Layer Normalization. This reduces the dimensionality while preserving the learned temporal features. | |
| ### 3. Self-Attention & Residual Connection | |
| The 512-dimensional features are passed through an 8-head Multi-Head Self-Attention mechanism. This allows the model to capture long-range temporal dependencies across the audio sequence. Finally, a residual connection adds the attention output back to the input, followed by a final Layer Normalization. | |
| ## Purpose | |
| By transforming the 768-dim BEATs embeddings into 512-dim primitives, this module creates a standardized, dense representation that the downstream Causal Graph Network and Memory modules can efficiently process. | |