SIBYLLE: a system for alternative communication adapting to the context and its user
Tonio Wandmacher, Jean-Yves Antoine, Franck Poirier · 2007 · Proceedings of the 9th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '07) · doi:10.1145/1296843.1296878
Summary
This paper describes the latest version of SIBYLLE, an augmentative and alternative communication (AAC) system developed at the Université François Rabelais and the Université Européenne de Bretagne for users with severe motor and speech impairments — including cerebrally and physically disabled users and people living with locked-in syndrome. SIBYLLE comprises four modules: a single-switch physical input device (eye glimpse, breath sensor, etc.), an on-screen virtual keyboard with multiple sub-keypads (letters, numbers, punctuation, predicted words, pre-recorded "emergency" sentences), a text editor that interfaces with the Windows API to enter text into any application, and a text-to-speech module. Three selection modes are offered (mouse, line/column scan, linear scan), and a sliding "timing line" provides dynamic feedback so users can anticipate when the scan frame will advance. The technical heart of the paper is the prediction machinery: SibyLetter dynamically reorders the letter keypad using a 5-gram character model, while SibyWord predicts whole words using a 4-gram statistical language model that is interpolated with two adaptive components — a Dynamic User Model trained continuously on text the user composes, and a Latent Semantic Analysis model that biases predictions toward words semantically related to the current topic. Geometric interpolation with confidence-weighted coefficients combines the models. SIBYLLE supports French, German, and English, and earlier versions had been used at the Kerpape rehabilitation centre in Brittany since 2001.
Key findings
On a French newspaper test corpus, SibyLetter's dynamic 5-gram letter prediction places the desired character at average position 2.7, compared to 9.0 average shifts for a static line/column scan and 33.0 for a static linear QWERTY scan, while requiring only a single keystroke per character. SibyWord with the full SIBYWORD configuration (4-gram + Dynamic User Model + LSA) achieves keystroke saving rates above 50% across newspaper, scientific, prose, speech, and email corpora in French, peaking at 59.4% on newspaper text. The Dynamic User Model alone produces large gains (e.g. +8.2% on scientific text, +9.4% on speech), reaching a +2% ksr lift after only ~2,000 words of user training (roughly 10 hours of typing). LSA contributes a smaller but consistent +1.0–1.7%, and the gains stack with DUM. German performance trails French by 4–6%, attributed to the prevalence of compound words. At Kerpape, more than twenty patients have used the system over seven years, reporting faster text entry, longer comfortable working sessions (less physical fatigue), and significantly fewer orthographic and grammatical errors. The authors flag a notable cognitive cost: users frequently miss valid predictions because reading the prediction list while writing increases cognitive load, motivating planned shifts toward direct word completion or embedding predictions into the letter keypad itself.
Relevance
This paper is a useful reference for anyone designing or evaluating scanning AAC systems, single-switch text entry, or word-prediction components for assistive communication. The clearest practical takeaway is that adapting language models to the individual user — not just the topic — produces substantial keystroke savings within hours of use, which has implications for any text-entry tool intended for low-bandwidth input (eye gaze, switch scanning, head pointing). The honest discussion of cognitive load — that more predictions are not always better, because users struggle to read the list and compose simultaneously — remains a live design tension in modern AAC and predictive-text interfaces. Limitations include the reliance on theoretical keystroke saving rate rather than measured communication speed, the limited evaluation with users who have additional cognitive or language impairments, and the now-dated reliance on newspaper training corpora; current systems would substitute neural language models and conversational corpora.
Tags: augmentative and alternative communication · AAC · word prediction · virtual keyboard · single switch · scanning · latent semantic analysis · language model · keystroke saving rate · text-to-speech · locked-in syndrome · cerebral palsy · rehabilitation · multilingual accessibility · natural language processing