Add GenAID vs WhisAID architecture comparison

README.md

@@ -14,6 +14,46 @@ tags:
 
 GenAID is a Chinese accent encoder based on `facebook/wav2vec2-large-xlsr-53`. It produces a 64-dimensional accent embedding designed to reduce speaker information. The model recognizes nine labels: north, Sichuan, Guangdong, south, Henan, Shanghai, Wuhan, Tianjin, and Singapore.
 
+## GenAID and WhisAID Medium side by side
+
+The comparison excludes Whisper's text decoder because accent identification uses only its audio encoder.
+
+| Component | GenAID (this model) | WhisAID Medium |
+|---|---|---|
+| Model-native input | Raw 16 kHz waveform `[B, T]` | 80-bin Whisper log-Mel `[B, 80, 3000]` |
+| Backbone | Wav2Vec2 XLSR-53 Large | Whisper Medium audio encoder |
+| Audio frontend | 7 convolution layers, combined stride 320 | 2 convolution layers, second layer stride 2 |
+| Frontend output | `[B, T′, 512]` | `[B, 1500, 1024]` |
+| Feature projection | `512 → 1024` | None after the CNN |
+| Transformer blocks | 24 | 24 |
+| Transformer width | 1024 | 1024 |
+| Attention heads | 16 × 64 dimensions | 16 × 64 dimensions |
+| FFN per block | `1024 → 4096 → 1024` | `1024 → 4096 → 1024` |
+| Transformer output | `[B, T′, 1024]` | `[B, 1500, 1024]` |
+| Pooling | Attention-mask-aware temporal mean | Temporal mean |
+| Accent embedding | `1024 → 64 → 64` | `1024 → 256` |
+| Accent output | `64 → 9` | `256 → 9` |
+| Speaker disentanglement | `64 → 336`; uniform-distribution adversarial MSE | GRL + speaker-token attention, `256 → 336` |
+| Task-relevant parameters | **315,530,560** | 306,473,897, excluding text decoder |
+| Frozen parameters | 4,210,176 (7-layer XLSR CNN) | Audio encoder is evaluated under `torch.no_grad()` in the compared recipe |
+| Effectively updated parameters | **311,320,384** | 444,249 |
+
+`T′` is the downsampled XLSR sequence length. For a 10-second, 16 kHz waveform, `T=160,000` and `T′≈499`.
+
+### GenAID module details
+
+| Module | Layers | Input → output | Parameters | Training behavior |
+|---|---:|---|---:|---|
+| XLSR feature extractor | 7 CNN layers | `[B,T] → [B,T′,512]` | 4,210,176 | Frozen |
+| Feature projection + XLSR encoder | Projection + 24 Transformer blocks | `[B,T′,512] → [B,T′,1024]` | 311,228,544 | Updated |
+| Masked mean pooling | 1 | `[B,T′,1024] → [B,1024]` | 0 | No parameters |
+| Information bottleneck | 2 × Linear + GELU | `1024 → 64 → 64` | 69,760 | Updated |
+| Accent classifier | Linear, no bias | `64 → 9` | 576 | Updated |
+| Speaker adversarial classifier | Linear, no bias | `64 → 336` | 21,504 | Updated during training |
+| **Total** | | | **315,530,560** | |
+
+The 64-dimensional output of the second GELU is the reusable accent embedding. The speaker classifier is a training-only probe: its predicted speaker distribution is optimized toward a uniform distribution, discouraging the bottleneck from retaining speaker identity. The speaker branch is not required when extracting embeddings.
+
 ## Usage
 
 ```python