| # Épée v0.2 · Signer-Robustness Benchmark |
|
|
| **What it shows:** signer variation is the core problem in sign-language AI, and the fix is |
| *more native signers and more data*, not a bigger model. |
|
|
| The model is trained on the four signers in this release (ALPHA, BRAVO, CHARLIE, DELTA) and tested on |
| **two native Deaf signers held entirely outside this release** (no shared clips, no shared phrases), so |
| there is no train/test leakage. All signers are pseudonymized. |
|
|
|  |
|
|
| --- |
|
|
| ## Setup |
|
|
| | | | |
| |---|---| |
| | Model | small BiLSTM (1 layer, hidden 64, bidirectional, mean-pool → linear) | |
| | Input | keypoint sequences, hands + arms (48 pts, x/y), shoulder-normalized, resampled to 24 frames | |
| | Task | sign (gloss) classification on **24 glosses** shared by all four training signers and both held-out test signers | |
| | Train | 858 gloss segments from the four release signers | |
| | Test | two signers held entirely out of training (≈1,900 and ≈1,450 segments) | |
| | Chance | 4.2% (1 / 24) · **8 seeds** per configuration, mean ± std | |
| | Metrics | top-1 accuracy **and macro-F1** (the task is imbalanced; macro-F1 is the honest headline) | |
|
|
| --- |
|
|
| ## Result 1 · More signers → better on a brand-new signer |
|
|
| Tested on a held-out signer (Signer A), never seen in training: |
|
|
| | # training signers | Accuracy | Macro-F1 | |
| |---|---|---| |
| | 1 | 22% | 0.13 | |
| | 2 | 40% | 0.24 | |
| | 3 | 50% | 0.32 | |
| | **4** | **59%** | **0.38** | |
|
|
| Each added signer lifts accuracy on a stranger; the 4th signer alone adds **+9 points**. A signer the |
| model *has* seen scores 65% (ceiling); chance is 4.2%. Confirmed on a second held-out signer (Signer B): **54%**. |
|
|
| > *Honest note:* run-to-run std narrows from ±12% (1 signer) to ±8% (4 signers), but part of that is a |
| > counting effect (more single-signer combinations at N=1 than at N=4). We therefore lead with the |
| > **accuracy trend**, not the variance. |
|
|
| ## Result 2 · More data → keeps climbing, not saturated |
|
|
| Same held-out signer; training set grown by a clean random fraction of the four-signer pool: |
|
|
| | Training examples | 80 | 160 | 260 | 380 | 520 | 680 | 858 | |
| |---|---|---|---|---|---|---|---| |
| | Accuracy | 23% | 30% | 33% | 43% | 42% | 54% | **61%** | |
|
|
| The curve rises from 23% to 61% and is still climbing at the largest size. More data directly buys more |
| generalization to an unseen signer. |
|
|
| --- |
|
|
| ## Takeaway |
|
|
| > Trained on one Deaf signer, the recognizer gets **22%** on a signer it has never seen. Trained on four, |
| > it reaches **59%**, and the curve (in both signers and data) has not flattened. A bigger model does not |
| > close this gap; it is a property of the data. This is why CLERC builds the multi-signer data layer. |
|
|
| ## Caveats |
|
|
| - 24-gloss shared subset; the full vocabulary will differ, but the *shape* (more signers/data → up) should hold. |
| - Measures generalization across signer **bodies**, not across topics; cross-vocabulary generalization is a separate, harder test. |
| - "Not saturated" is supported by the signer curve and 7 data points, not proven asymptotically. |
| - Small BiLSTM by design; a bigger model does not close the 1-signer-vs-4-signer gap, which is a data property. |
| - The two held-out test signers come from the broader CLERC corpus and are not part of this public release. |
|
|
