Private ASR V3

Privacy-first transcription workspace for uploaded audio and video records.

This repository is intentionally scoped to the non-realtime part of the product:

• file ingestion
• record storage and retrieval
• transcript and speaker editing
• re-transcription and AI re-organization
• agent-friendly APIs and skills

Realtime browser recording and live WebSocket transcription are not part of this repository. Those capabilities stay archived in the previous codebase.

Project definition documents live under Project_Definitions.

The default deployment model is Docker Compose with separate api, web, and worker services.

Bootstrap Stack

The initial runnable scaffold keeps a few practical choices from the archived repo:

• Express API
• JWT authentication
• SQLite for bootstrap persistence
• local volume-backed file storage

What changed:

• no realtime recording
• no WebSocket proxy
• no chunk upload or browser recovery logic
• background processing moved to a dedicated worker

Quick Start

cp .env.example .env
# edit .env with your local secrets and provider endpoints
docker compose build
docker compose up

Then open http://localhost:8080.

Runtime configuration is loaded from .env. A sanitized template lives in .env.example. Docker Compose passes .env into both api, worker, and web, and the services also load /app/.env directly for non-Docker local runs.

For the web service, the frontend backend endpoint is injected at container start through API_BASE_URL. If left blank, the UI falls back to http(s)://<current-host>:3000/api/v1.

Related Projects

• ASR_API_V2: the self-hosted ASR service used by this workspace for transcription

This repository does not vendor the ASR service as a submodule. Private-ASR-V3 treats transcription as an external dependency and connects to it through ASR_API_URL.

External Service Configuration

Configure external providers in .env by following the field descriptions in .env.example.

Web App Backend Endpoint

Set API_BASE_URL to the full browser-facing app backend base URL, for example:

• https://asrapi.example.com/api/v1

If API_BASE_URL is left empty, the frontend falls back to the current host on port 3000, matching the old repo's split deployment convention.

For production, set API_BASE_URL explicitly. Also include your frontend origin in FRONTEND_URL so the API CORS allowlist matches the deployed site.

ASR

Point the workspace to your deployed ASR service:

• ASR_API_URL
• ASR_LANGUAGE
• ASR_TIMEOUT

The recommended ASR implementation is ASR_API_V2.

Speaker diarization is enabled by default in the worker. Hotwords are request-level input from the UI/API and are forwarded to the ASR provider when present, so neither of them is configured through .env.

LLM

LLM-based summary generation is configured the same way through .env:

• LLM_API_BASE
• LLM_API_KEY
• LLM_MODEL
• LLM_TIMEOUT
• LLM_BATCH_MAX_LENGTH
• LLM_BATCH_DELAY_MS

In practice, just copy .env.example to .env and fill in the LLM endpoint and key for your provider.

LLM Summary Notes

The worker supports long-transcript summary batching.

• LLM_BATCH_MAX_LENGTH: maximum characters sent to the LLM in one summary request
• LLM_BATCH_DELAY_MS: delay between batch requests to reduce provider throttling

If a transcript is longer than LLM_BATCH_MAX_LENGTH, the worker splits it into multiple batches, summarizes each batch, and then attempts a final consolidation pass.