VoiceBridge

I built VoiceBridge for my uncle while preparing to leave for military service. His speech is difficult for our family to understand; often, we cannot make out the key word. VoiceBridge aims to help him take part in everyday conversations: answering a question, asking for help, or starting a different topic. During this hackathon, I built the conversation flow, family confirmation tools, and "Find my words" interaction. The app listens to the conversation partner. Muse Voice Transcribe turns the question into text, and Muse Spark 1.3 proposes up to four short replies using the question, recent conversation, and relevant family-confirmed information. Tapping a sentence requests device text-to-speech for that exact text. The person can reject suggestions, use a fixed response, stop, or open "Find my words" to explore another expression. Find offers sentence choices or optional Korean-initial and English-alphabet input. Letter clues remain exactly as entered, so a family member can interpret them and ask the person to confirm. The app does not automatically turn a letter into a sentence. A separate family page lets relatives correct a misheard question, review information about familiar people and preferences, and save useful expressions. Spark asks one interview question at a time to help families provide background. Draft notes require family review and PIN confirmation. Choosing or rejecting a reply never silently changes a lasting preference. The design keeps generated options, family interpretations, and the person's decision distinct. The app screens candidates and discards outdated responses. Fixed expressions remain available without waiting for a new model response. I tried VoiceBridge with my uncle in two bedside sessions on the same day. I did not observe unassisted app speech in the first session. Some replies in the second session followed his own taps, sometimes after family prompts; other turns still needed hands-on assistance. Fuller source footage includes a family-assisted letter-clue exchange with verbal confirmation of a body location. The short demo does not show his answer to its later, more specific confirmation question. His taps clustered near the lower screen, so a tap alone did not establish intended meaning. These are early observations with one person, not evidence of independent everyday use or clinical effectiveness. Personalization was tested separately with synthetic family information. Next, I want to compare touch with a joystick and selection button, or large external switches. Moving the highlight would not speak a reply. I also want Find and reviewed family information to help him introduce his own topics. These are untested proposals. VoiceBridge is an assistive communication aid, not a medical device. It is not designed to diagnose conditions or give medication advice.