Baseline on Bukva recognizes Russian fingerspelling letters with 83.6% top-1 accuracy

At WSLP-2026 (Workshop on Sign Language Processing), held as part of EMNLP in Budapest, researchers presented three papers on fingerspelling recognition for Russian Sign Language (RSL). The headline number comes from the Bukva dataset: a baseline built on the lightweight TSM and MobileNetV2 models reached 83.6% top-1 accuracy, meaning the correct letter was the model's first answer in 83.6% of cases. The other two papers are HandReader, which scored 75.6 on the ChicagoFSWild+ benchmark, and the Znaki dataset with 37 252 videos. The authors could not attend in person, so a colleague brought the posters and answered visitors' questions himself. The authors promise to release the code publicly.

Bukva: an open RSL alphabet and a trainer that runs on an ordinary CPU

Bukva is the first open dataset for recognizing the RSL fingerspelling alphabet. It contains 3757 videos, 33 letters, 8 of which are dynamic (signed with a hand movement rather than a static pose), and 155 native signers.

The authors built the baseline from lightweight components: TSM (Temporal Shift Module, which shifts features along the time axis) and MobileNetV2. The result is 83.6% top-1 accuracy. The model runs in real time on a single core of a mid-range processor and needs no GPU. On this basis, the team built a demo trainer for learning fingerspelling. For educational apps on ordinary laptops and phones, that light footprint matters more than record accuracy.

HandReader: three architectures and a result on ChicagoFSWild+

The second paper, HandReader, describes three architectures for fingerspelling recognition:

• HandReader_RGB works with ordinary video and uses a new module called TSAM. It processes clips of different lengths without padding or cropping, so it neither fills short recordings with blank frames nor cuts long ones;
• HandReader_KP relies on a hand keypoint encoder rather than on pixels;
• HandReader_RGB+KP combines both modalities.

The family claims SOTA on the foreign ChicagoFSWild+ benchmark (75.6). ChicagoFSWild and ChicagoFSWild+ are sets of American fingerspelling videos shot in everyday conditions rather than in a studio. On ChicagoFSWild, the model posted the best result among models trained only on their own pretraining, that is, without external pretrained weights. The post does not specify which metric the 75.6 figure refers to, so the exact comparison conditions should be checked in the paper on arXiv.

Znaki: continuous fingerspelling and proper names

The third paper moves from individual letters to whole sequences. Znaki is the first open dataset of continuous fingerspelling for RSL. It contains 37 252 videos with a total duration of 35.4 hours and 1593 unique sequences. These are proper names from 16 categories. The dataset was recorded by 68 RSL signers, including certified interpreters.

The data was collected through crowdsourcing: contributors were selected in two stages, and experts checked the results.

As baselines, the authors used models adapted for American Sign Language (ASL). They reach up to 0.932 on the 1−CER metric (one minus the character error rate) and 0.596 on sequence accuracy, the share of sequences recognized entirely without errors. In other words, the models guess most characters correctly, but they read a whole sequence without a single error in only slightly more than half of cases.

What this means

Together, the three papers give Russian-language research open data for both individual letters and continuous fingerspelling, along with baselines for comparison. The results have limits: fingerspelling is only one part of RSL and does not replace recognition of full signs and grammar, and all metrics were obtained on specific datasets. Once the authors release the promised code, it will be easier for developers to verify the reported numbers and build their own apps for learning and translation.