KhalidIbrahem/somali-music-archive

# Somali Music AI Preservation Platform **The first AI-powered archive of Somali traditional music.** Somali music — *heello*, *qaraami*, *dhaanto*, *buraanbur*, and the sung poetry of *gabay* — is a living oral tradition with no representation in any major music dataset or MIR corpus. This platform records, preserves, annotates, and teaches it: a mobile field-recording app, an AI enrichment pipeline that transcribes lyrics and measures the tradition's microtonality, a public archive with search and lessons, and a documented Research API for the ethnomusicology community. Built by **Khalid Ibrahim** — Somali-American AI engineer, Minneapolis MN. ## What makes this technically interesting - **A tonal system, measured for the first time.** CREPE pitch tracks are mapped onto a Somali pentatonic reference and each frame's **deviation from Western equal temperament is quantified in cents** (`apps/ai-service/services/scale.py` — deliberately dependency-free and unit-tested). The reference frequencies are calibrated against master performers, not asserted from Western theory. This is the core contribution of the accompanying ISMIR paper draft (`docs/ISMIR_DRAFT.md`). - **Sung text is not speech.** Whisper hallucinates over melisma; every transcript segment is gated by hallucination heuristics and flagged `is_singing`, stored as advisory rather than ground truth. - **A 605-recording research corpus** from Harvard's Archive of World Music is processed end-to-end by `apps/ai-service/notebooks/harvard_pipeline.ipynb` (denoise → source-separate → transcribe → pitch-map → dataset), then ingested via `npm run ingest:harvard` — landing in the moderation queue, never auto-published. ## Architecture at a glance ``` apps/mobile React Native (Expo SDK 57) — recording, archive, lessons, offline apps/web Next.js 15 — marketing site + admin dashboard (moderation, orgs) apps/api Node 20 + Express — auth (JWT+RBAC), recordings, search, Stripe, collections/commen …