The only supported public deployment posture is production mode:

PRODUCTION_MODE=1
JWT_SECRET=<long random token>
CORS_ALLOW_ORIGINS=https://your-frontend.example
MAX_AUDIO_SECONDS=<positive integer>
RATE_LIMIT_REQUESTS=60
RATE_LIMIT_WINDOW_SECONDS=60

In production mode the API:

• fails startup if JWT_SECRET is missing,
• fails startup if CORS_ALLOW_ORIGINS is empty or contains *,
• fails startup if MAX_AUDIO_SECONDS <= 0,
• hides internal details for most 5xx errors,
• enforces upload size via MAX_UPLOAD_MB,
• enforces audio duration via MAX_AUDIO_SECONDS when duration can be read,
• enables in-memory rate limiting for /v1/... endpoints.

Development mode may allow public endpoints when JWT_SECRET is empty. Do not use development mode on an internet-facing host.

Report suspected vulnerabilities privately to the repository owner or maintainer. Include:

• affected endpoint or script,
• steps to reproduce,
• impact,
• whether secrets, personal data, audio, transcripts, model files, or database records may be exposed,
• suggested fix if known.

Do not open public issues containing exploit details, credentials, personal data, private audio, or database dumps.

Never commit:

• .env
• OpenAI API keys,
• database URLs with real passwords,
• model download tokens,
• private service URLs,
• production hostnames if they are sensitive,
• database dumps,
• logs,
• exported terms containing private content.

Use environment variables or secret mounts. Prefer OPENAI_API_KEY_FILE where the deployment platform supports file-based secrets.

The current API authentication is intentionally simple:

Authorization: Bearer <JWT_SECRET>

It is not full JWT validation. The token string is compared directly with JWT_SECRET.

For multi-user public deployments, replace this with a real auth layer:

• signed JWT validation,
• key rotation,
• per-user authorization,
• audit logging,
• token revocation.

Production deployments must set explicit origins:

CORS_ALLOW_ORIGINS=https://lemko.tools,https://www.lemko.tools

Avoid * in production. If browser credentials are needed, configure a narrow origin list and set CORS_ALLOW_CREDENTIALS=1 only when required.

Rate limiting is in-memory and process-local. It is suitable as a basic protection, not as the only public abuse control.

Configuration:

RATE_LIMIT_REQUESTS=60
RATE_LIMIT_WINDOW_SECONDS=60
TRUST_PROXY_HEADERS=1

If the app is behind Caddy or another trusted proxy, set TRUST_PROXY_HEADERS=1 so X-Forwarded-For is used. Do not trust forwarded headers when clients can connect directly to the app.

For scaled deployments with multiple API replicas, use proxy-level or centralized rate limiting.

The API accepts:

• audio/wav
• audio/x-wav
• audio/flac
• audio/mpeg
• application/octet-stream

Controls:

• MAX_UPLOAD_MB limits request body size after read.
• MAX_AUDIO_SECONDS rejects audio if torchaudio.info() can determine the duration and it exceeds the limit.

Recommended additional hardening:

• enforce body size at reverse proxy level,
• run antivirus or media validation for untrusted uploads,
• store uploaded audio on encrypted storage,
• delete source uploads according to retention policy.

Production mode suppresses internal 5xx details in API responses. Internal details may still be written to logs for operators.

Do not return:

• filesystem paths,
• database connection strings,
• model paths revealing internal layout,
• stack traces,
• raw upstream API errors,
• private term/source content in error details.

This repo includes:

• .github/dependabot.yml
• .github/workflows/pip-audit.yml

Run locally:

python3 -m pip install pip-audit
pip-audit -r requirements-prod.txt

Dependency results change over time. Treat every production build as requiring a fresh audit.

As of the user request dated 2026-05-25, known advisory concern was raised for several pinned dependencies including torch==2.3.1, python-multipart==0.0.20, starlette==0.47.3, nemo-toolkit==2.4.0, transformers==4.56.0, and nltk==3.9.1. Verify current fixed versions before upgrading because PyTorch, NeMo, and Transformers compatibility can be tightly coupled.

The local audit result from this work is summarized in docs/DEPENDENCY_AUDIT.md.

TTS/voice cloning can be abused for impersonation. Public deployments should:

• require explicit consent for every speaker reference set,
• restrict who may generate TTS,
• watermark or label generated speech where appropriate,
• log abuse reports,
• support deletion and withdrawal of voice consent,
• block deceptive or harmful impersonation use cases.

See MODELS.md and PRIVACY.md.

• PRODUCTION_MODE=1.
• Non-empty JWT_SECRET.
• Explicit CORS_ALLOW_ORIGINS.
• Reverse proxy body size limits.
• HTTPS only at the public edge.
• No direct public access to Postgres, Qdrant, or Uvicorn.
• Encrypted backups for database and logs.
• Regular dependency audit.
• Retention/deletion job for logs, source audio, and transcripts.
• Separate secrets from image and repository.