Fix fabricated citation; credit actual authors

README.md

@@ -39,35 +39,56 @@ is original to this repo.
 
 | File | Size | Description |
 |---|---|---|
-| `baseline.pt` | 11 MB | DNS5-trained baseline (val_loss 0.0213, ERLE 14.9 dB). |
-| `finetune.pt` | 11 MB | AEC fine-tune (1 epoch on AEC Challenge synthetic set). |
-| `localvqe-baseline.gguf` | 5 MB | GGML F32 export of `baseline.pt` (BN-folded, DCT weights embedded). |
-| `localvqe-finetune.gguf` | 5 MB | GGML F32 export of `finetune.pt`. |
+| `localvqe-v1.pt` | 11 MB | PyTorch checkpoint — DNS5 pre-training + AEC Challenge fine-tune. |
+| `localvqe-v1-f32.gguf` | 5 MB | GGML F32 export (BN-folded, DCT weights embedded). |
 
 ## Usage (GGML / C++ / Go)
 
 ```bash
 cd ggml && cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
-./build/bin/localvqe localvqe-baseline.gguf \
+./build/bin/localvqe localvqe-v1-f32.gguf \
     --in-wav mic.wav ref.wav --out-wav enhanced.wav
 ```
 
 Per-frame wall time on Zen4 (24 threads): ~1.66 ms (9.6× realtime at
 16 kHz / 256-sample hop).
 
+## Validation
+
+Best checkpoint of the published v1 F32 weights (`localvqe-v1-f32.gguf`), evaluated
+on a 1 000-clip validation split mixing DNS5-synthesised scenes and ICASSP
+AEC Challenge synthetic data; AECMOS over a 100-clip sub-sample.
+
+| Metric | Overall | Single-talk far-end | Double-talk |
+|---|---:|---:|---:|
+| ERLE | — | **+52.2 dB** | — |
+| AECMOS echo (↑) | 4.36 | 4.46 | 4.33 |
+| AECMOS degradation (↑) | 4.83 | 5.00 | 4.78 |
+
+ERLE is only meaningful on single-talk far-end (mic = pure echo); overall
+and double-talk ERLE are confounded by near-end speech and are omitted.
+
+DNSMOS OVRL is deliberately omitted: our P.808 scores (~2.0) contradict
+informal listening of the single-talk far-end output, which is audibly
+near-silent. We suspect our DNSMOS invocation is miscalibrated for AEC
+outputs — particularly near-silent clips, which are out of distribution
+for a speech-quality predictor — and would rather withhold the numbers
+than publish misleading ones.
+
 ## Architecture
 
 | Component | Value |
 |---|---|
 | Sample rate | 16 kHz |
 | Analysis basis | DCT-II (Conv1d filterbank, 512 filters, stride 256, frozen) |
-| Mic encoder | 5 blocks: 2 → 64 → 128 → 128 → 128 → 128 |
-| Far-end encoder | 2 blocks: 2 → 32 → 128 |
+| Mic encoder | 5 blocks: 2 → 32 → 40 → 40 → 40 → 40 |
+| Far-end encoder | 2 blocks: 2 → 32 → 40 |
 | AlignBlock | Cross-attention soft delay, d_max=32 (320 ms), h=32 |
-| Bottleneck | S4D diagonal state-space |
-| Decoder | 5 sub-pixel conv + BN blocks: 128 → 128 → 128 → 64 → 64 |
-| CCM | 27-ch → 3x3 complex convolving mask (real-valued arithmetic) |
-| Parameters | ~8.0 M |
+| Bottleneck | S4D diagonal state-space, hidden 162 |
+| Decoder | 5 sub-pixel conv + BN blocks, mirroring encoder |
+| CCM | 27-ch → 3×3 complex convolving mask (real-valued arithmetic) |
+| Kernel | (4, 4) time × freq, causal padding |
+| Parameters | ~0.9 M |
 
 ## Licence