ABSTRACT

SpeakAgain version 1.0 (described in a companion paper) demonstrated the technical feasibility of a free, open-source web platform combining a deterministic aphasia severity classifier, a large-language-model sentence completion engine, and Google text-to-speech (gTTS) synthesis for augmentative communication and home-based rehabilitation. Real-world use of version 1.0 surfaced six concrete deficiencies: per-session state loss, no authentication, no cross-session family visibility, indiscriminate caregiver email volume, partial UI localisation, and non-native voice output for Nigerian languages. This paper presents SpeakAgain version 2.0, an architectural redesign that addresses each deficiency. The release introduces a server-side shared patient store that persists profile, assessment outcome, experience-point totals, streak history, recovered-word catalogue, and family registry across logins and across browser sessions; dual authentication via username-and-password (PBKDF2-HMAC-SHA256, 200,000 iterations, constant-time comparison) and Streamlit-native Google OpenID Connect (OIDC) with email-based account linking that prevents duplicate accounts from a single user; a read-only family dashboard reachable through a one-shot 24-hour invite code with no account creation required by family members; a deterministic event-dispatch router that routes crisis events (pain, falls, breathing difficulty) to email and routes all routine activity to the dashboard, eliminating the notification-flooding observed in version 1.0 deployments; full user-interface localisation through a centralised translation function covering eighty-two namespaced keys across nine languages; five fully-playable mini-games with emoji-based pictographic anchors; and a hybrid voice synthesis layer that routes the four Nigerian languages (Yoruba, Igbo, Hausa, Nigerian Pidgin) to YarnGPT for native voice output and routes the other five languages (English, Spanish, French, Portuguese, Arabic) to gTTS, with a session-scoped circuit breaker to bound latency under vendor failure. We describe the architecture, the event-dispatch logic, the invite-code lifecycle, the YarnGPT integration with its fallback policy, the security properties, and the relationship of version 2.0 to the priorities surfaced in the 2025 systematic review of self-administered digital aphasia therapy (Ericson et al., 2025). The source code is released under a permissive licence at github.com.

Keywords: aphasia, stroke rehabilitation, augmentative communication, OpenID Connect, multilingual text-to-speech, family-centered care, low-resource settings, Nigeria