OpenTranscribe — System Architecture

OpenTranscribe — System Architecture An architecture diagram generated by Archify. User · Browser · Architecture component User Browser MinIO · media/object storage · Architecture component MinIO media/object storage SvelteKit SPA · frontend/src · Architecture component SvelteKit SPA frontend/src Nginx · reverse proxy · Architecture component Nginx reverse proxy OpenSearch · hybrid/neural search · Architecture component OpenSearch hybrid/neural search FastAPI Backend · backend/app/api · Architecture component FastAPI Backend backend/app/api Redis · broker + cache · Architecture component Redis broker + cache PostgreSQL · primary DB · Architecture component PostgreSQL primary DB Celery Workers · pipeline plan per file · Architecture component Celery Workers pipeline plan per file LLM Provider · local or remote · Architecture component LLM Provider local or remote Cloud speech APIs · required on lite · Architecture component Cloud speech APIs required on lite HTTPS presigned GET/PUT fetch (cookie auth) proxy_pass search query SQL enqueue task dispatch write results index transcript chat / summarize read/write media audio, cloud steps only

Web tier

  • • SvelteKit SPA, httpOnly-cookie auth, no JS-readable token
  • • Nginx terminates TLS, proxies /api and /ws
  • • Frontend media reads/writes go via presigned MinIO URLs — the backend only generates the presign URL string, Celery workers do the actual streaming

Backend + async pipeline

  • • FastAPI routers under backend/app/api
  • • Celery queues gpu, cpu, cpu-transcribe, cloud-asr, nlp, embedding, redaction, utility, download; the pipeline plan picks the transcription and speaker queues per file
  • • Transcription: local faster-whisper or a cloud vendor; speakers: this server, pyannote.ai or a vendor; voiceprints always on this server

Data + retrieval

  • • Postgres is system of record
  • • OpenSearch does hybrid/neural transcript search
  • • MinIO holds original media + derived artifacts
  • • Lite image: no local transcription or speaker detection, so the cloud vendors become required; voiceprints still run here on the CPU