Fluent: An AI Augmented Writing Tool for People who Stutter
Bhavya Ghai, Klaus Mueller · 2021 · The 23rd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2021) · doi:10.1145/3441852.3471211
Summary
This paper presents Fluent, a novel AI-powered writing tool designed to help people who stutter (PWS) prepare scripts and written content that they can deliver more fluently. Over 70 million people worldwide stutter, and a common coping strategy is word substitution — replacing words likely to trigger stuttering with easier-to-pronounce alternatives. However, this mental process of anticipating difficult words and finding substitutes in real time during speech is cognitively taxing and stressful. Fluent automates this process by using phonetic embeddings from the CMU Pronouncing Dictionary to represent words based on their constituent phonemes, then training a binary SVM classifier to distinguish between words a specific individual finds easy versus difficult to pronounce. The system uses active learning with uncertainty sampling to efficiently learn each user's personalized phonetic difficulty patterns from minimal feedback. Users provide initial seed words (at least 5 easy and 5 difficult words), and the classifier highlights potentially difficult words in blue as the user types. Hovering over a highlighted word reveals a dropdown of semantically similar but phonetically easier alternatives sourced from the DataMuse API, which combines multiple dictionaries, WordNet, and Word2Vec embeddings. The system gathers both implicit feedback (whether users accept or ignore suggestions) and explicit feedback (directly labeling words as easy or difficult), continuously retraining the classifier to improve personalization.
Key findings
Evaluation using 2,467 TED talk transcripts (57,000 unique words) across 10 simulated user profiles showed that Fluent achieves mean accuracy over 80% in identifying trigger words within just 20 interactions, and performance continues improving with additional feedback. User profiles were modeled from real stuttering patterns gathered from online communities and one author's personal experience, each defined by specific phonetic patterns (e.g., difficulty with consonant-r combinations, words starting with 'st' or 'fl', words containing 'sc' or 'ch'). Explicit feedback proved more effective than implicit feedback at accelerating the classifier's learning, as a single explicit interaction can add two data points (confirming a word as difficult and adding an alternative as easy). The confidence threshold for highlighting words significantly affects performance — lower thresholds catch more trigger words but produce more false positives. The system correctly handles proper nouns, names, and places by identifying them via Named Entity Recognition and flagging them without offering substitutions. The tool is implemented as a Flask web application with a Summernote-based rich text editor, designed to be cross-platform accessible without specialized software.
Relevance
Fluent represents an innovative application of AI to an underserved accessibility need — there has been virtually no prior work on assistive writing technology specifically for people who stutter. While most AI-stuttering research focuses on detection and classification of disfluencies in speech, Fluent takes the opposite approach: helping users avoid stuttering events proactively through script preparation. For accessibility practitioners, this tool illustrates how personalization through active learning can address highly individual accessibility needs — each person who stutters has a unique phonetic profile. The authors are transparent about limitations: word substitution is a coping mechanism rather than a treatment, concealing stuttering can negatively affect self-esteem, and the tool works best for planned speech scenarios like presentations or talks rather than spontaneous conversation. The open-source availability of Fluent and its potential extension to other speech conditions (lisp, non-native speakers) and integration into mainstream writing tools like MS Word makes it a practical contribution to assistive technology.
Tags: stuttering · speech disorders · machine learning · active learning · natural language processing · writing tools · assistive technology · word substitution