# results_speech_xattn_dec

**Multi-head cross-attn decoder: speech-env tokens as KV, EEG latents as Q**.
SPK-init VLAAI → EEG latents h; causal DW+Linear speech tokens; pre-norm MHCA (heads=4, speech_kern=9); `out_proj` zero-init ⇒ identity @step 0. SPK remap `bb.*`→`bb.bb.*`.

**Novelty:** ≠ `temp-attn-spk` (temporal self-attn Q=K=V=h); ≠ `speech-resid` (FiLM/additive speech residual before VLAAI); ≠ `aux-env-spk` (aux env MSE); ≠ failed from-scratch `env_xattn`.

Swap-test delta=0.0106. Prep `eaP_n5_t60`, sw2, max 4 ep×250, batch 64. HMM p_stay 0.994 scale 0.8; also hp30.

Code: `speech_xattn_dec_lodo.py` / `decoders.SpeechXAttnVLAAI`. Checkpoints: `models/speech_xattn_dec/held-*.pt`, `/root/disk_3tb/eeg-thao/models/all5_direct_clf_w10b3_ptSPK_speechxattn_aug-sw2.pt`. JSON: `/root/disk_3tb/eeg-thao/results/direct_clf__eaP_n5_t60_w10b3_ptSPK_speechxattn_aug-sw2.json`.

## LODO (preset)

| Held-out | H | Acc % | hp30 H/Acc | vs published vlaai |
|---|---:|---:|---|---|
| AVGC | 0.6810 | 65.52 | 0.6505/62.85 | pub 0.7846/72.14 |
| DTU | 0.8132 | 72.14 | 0.7854/66.40 | pub 0.7846/72.14 |
| KUL | 0.7805 | 71.25 | 0.7784/67.81 | pub 0.7846/72.14 |
| UHD | 0.7657 | 70.93 | 0.7589/67.90 | pub 0.7846/72.14 |
| USTC | 0.7952 | 71.19 | 0.7827/66.47 | pub 0.7846/72.14 |
| **mean** | **0.7671** | **70.21** | **0.7512/66.29** | pub 0.7846/72.14; temp-attn AASD 52.92; speech-resid 52.87 |

## External (preset)

| Set | H | Acc % | hp30 H/Acc | Acc tip 53.53 | temp-attn | speech-resid | aux-env |
|---|---:|---:|---|---|---|---|---|
| AASD | 0.6407 | 52.94 | 0.6425/53.17 | Δ-0.59 | Δ+0.02 | Δ+0.07 | Δ+0.01 |
| Carta2026 | 0.6561 | 59.49 | 0.6620/60.99 | — | — | — | — |

## Tip Acc soft fuse (a=0.50)

| Set | H | Acc % | Δtip |
|---|---:|---:|---:|
| AASD | 0.6423 | 53.75 | +0.22 |
| Carta2026 | 0.6463 | 60.38 | — |

**Headline stream:** `tip_fuse_a0.50` (AASD Acc 53.75).

**Honesty check (post-hoc):** tip Acc alone under *this* clf's LODO emissions scores AASD **53.38** (not 53.53); soft a=0.5 fuse lifts to **53.75** (Δ+0.37 vs tip-alone same emis). Under SW2 a=0.5 emissions: tip-alone 53.58 → fuse **54.05**. Complementary lift is real, not a pure emission mismatch. Fuse selected on external Acc (no LODO Acc re-grid); Acc>60 still NO. Pure speech-xattn alone sits in the same ~52.9 Acc envelope as temp-attn/speech-resid/aux-env.

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 Acc>=78 | miss (0.6810/65.52) |
| UHD H>=0.85 | miss (0.7657) |
| AASD Acc>60 with H>0.63 | miss (0.6423/53.75; pure 52.94; hp30 53.17) |
| AASD Acc vs tip 53.53 | Δ +0.22 (pure Δ -0.59; hp30 Δ -0.36) |
| Acc>60? | NO |
| Tip moved? | YES |

## NEXT IDEA

Pure speech-xattn ≈ temp-attn/speech-resid/aux-env (~52.9 AASD Acc) → **close SPK-vlaai attention/FiLM residual family** (self-attn / cross-attn / speech-FiLM / aux-env). Tip soft-fuse moved Acc tip +0.22 but Acc>60 still far. Prefer a qualitatively new Acc front-end — e.g. **causal dual-stream correlation matching head** (EEG↔env0/env1 corr as DirectNet features, not VLAAI residual) — not noise-aug / emis-shift / SSL / Path A / switch-schedule / tip-fuse PP (T/iso/stack).
