ReVoice
“Offline captions, speaker diarization, and conversation intelligence powered entirely by on-device AI.”
Traditional speech transcription and meeting summarizers stream private audio to third-party cloud APIs, incurring ongoing metered charges and presenting critical privacy risks for confidential business discussions, classrooms, and sensitive face-to-face meetings.
Co-developed ReVoice as part of Team LarpCode, where I focused on Android native development alongside the team's Flutter UI and iOS on-device AI engine. The system runs lightweight on-device models (Moonshine Voice and Qwen2.5 via llama.cpp) to deliver zero-cloud transcription, speaker corrections, structured clean transcripts, and grounded conversation Q&A.
Event Pipeline & Data Flow
Microphone audio stream → native on-device audio buffer → Moonshine Voice STT → Flutter transcript UI → Qwen2.5 on-device LLM via llama.cpp → local session summary & Q&A.Captures microphone audio locally on device; stream remains strictly within native processing layer with zero server uploads.
On-device Moonshine Voice model produces live streaming English captions with partial and finalized line detection.
Speaker labels, timestamps, and rule-based filler cleanup improve readability while preserving original recognized words.
Local Qwen2.5-1.5B via llama.cpp generates post-session executive summaries and answers questions with transcript citation references.