Layout-aware gesture recognition
Pass normalized x/y coordinates and timestamps with the key positions for your keyboard. The encoder is required; layout-specific refinement can be added where a matching decoder exists.
Downloadable swipe-recognition models and a C++ engine for turning keyboard gestures into word candidates.
FUTO Swipe is a developer release of trained models and recognition algorithms for gesture typing. Supply a swipe path, keyboard layout and dictionary, and its C++ engine returns predicted words. The models can run on a device, allowing developers to build swipe input into their own software.
FUTO developed the technology for FUTO Keyboard and also distributes it for other integrations. The encoder supports supplied keyboard layouts; the optional released decoder is English/QWERTY-specific, and the ContextLM is English-only. The separate research dataset contains one million English QWERTY swipes.
Pass normalized x/y coordinates and timestamps with the key positions for your keyboard. The encoder is required; layout-specific refinement can be added where a matching decoder exists.
The C++ library handles model inference and beam search against your dictionary, producing candidate words instead of leaving you with raw character scores.
The optional English ContextLM uses preceding words to improve candidate ranking and can operate with or without the layout-specific decoder.
Build the library for desktop Linux or Android. Its API lets you change layouts, dictionaries and optional models during use.
Integrate the recognition engine into your input interface, provide matching layout coordinates and vocabulary, and display candidate words to the user.
Evaluate swipe input on a trackpad or a custom virtual keyboard using your own layout and dictionary.
Use the released English QWERTY dataset to compare decoding approaches, while keeping benchmark results separate from performance in your own interface.
No. FUTO Keyboard uses FUTO Swipe. This release provides the models, inference library and research resources for developers to integrate into other software.
You can supply a layout and dictionary to the general encoder. The currently released decoder is English/QWERTY-only, and the ContextLM is English-only; they cannot be assumed to refine other combinations.
Weights are royalty-free under the FUTO Model Weights License 1.0, including commercial use with visible FUTO Swipe attribution for end users. Other use and distribution conditions also apply. The inference library is GPLv3 or later, and the dataset is MIT-licensed; each artifact has its own license.
The documented SwipeEngine class is not thread-safe. Protect it from concurrent use rather than assuming parallel calls are safe.
Os dados de tráfego são apenas para referência.