Suara: Patient Interviews for Family Memory

Suara explores patient family-memory interviews using AssemblyAI's adaptive baseline, an experimental patient configuration and a Streaming prototype. Patient settings require an exact provider echo; benefit is unproven. The live site has Bahasa Malaysia, English and Simplified Chinese interface copy. Its synthetic walkthrough now plays generated narration, follows timed excerpts, changes playback speed and exports selected, reviewed fictional quotes. These prepared transcripts are not live ASR. Managed speech is configured for English/Mandarin input with English replies; Malay live speech and code-switching accuracy are not claimed. Human recording remains on hold pending provider data controls, participant consent, restricted storage and independent review. Managed sessions are recorded by the provider; browser reset does not delete those recordings. Suara's token routes receive no audio/transcripts. Optional AI questions are default-off and can forward at most two final-answer excerpts through AssemblyAI LLM Gateway and its model provider. On 29 September, 716 automated tests passed; 15 credential-free hosted checks passed (static/health only). A paid public-data pilot completed 36 transcriptions of 12 CC-BY-4.0 FLEURS clips across three configurations. Context hints worsened Mandarin development errors and were not promoted. A tiny held-out subset is pipeline evidence, not a general accuracy claim; no model weights were trained. A separate synthetic English managed interview passed source-anchor, question-order and clean-end checks, with reply audio received. No physical microphone or human-heard response was verified. Reviewed quote arrangement is implemented; an AI-authored memoir, older-adult benefit and production readiness remain unproven. The public MIT-licensed source, synthetic walkthrough, video and deck are available for judges to inspect.