Beyond Transcript Alignment
Collection
Frozen-frozen speech-LLM adapters + counterfactual training. Code: github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress • 8 items • Updated
BTA — Stage 4 Control A (text-only adapter)
3-seed falsification control: replaces the WavLM audio path with a frozen 4096→1024 linear projection P fit on 500 Stress-17K samples (numpy lstsq, train MSE 0.090, train cos-sim 0.999). All other R1.8 hyperparameters are bit-identical.
Pre-registered PASS gate: MLP-2 cohort ≤ 0.31 (= K_T floor 0.290 + 0.02). Result: 0.2122 ≤ 0.31 → PASS, gap to R1.8 cohort = -0.094.
Files: A_textK_seed{1234,1235,1236}.pt (~336 MB each)
Reported metrics:
| Seed | Probe-K linear | Probe-K MLP-2 |
|---|---|---|
| 1234 | 0.2276 | 0.2125 |
| 1235 | 0.2393 | 0.1916 |
| 1236 | 0.2207 | 0.2323 |
| mean ± σ | 0.2292 ± 0.0094 | 0.2122 ± 0.0203 |
Code / paper: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress
License: CC-BY-NC-4.0.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support