aisrt 1.1.0

Hardware-aware, concurrent pipeline for subtitle generation.

🎬 Ultimate SRT Generator

Hardware-aware, zero-disk, highly concurrent AI pipeline for mass-generating broadcast-quality subtitles.

---

🌟 Overview

The Ultimate SRT Generator is a production-grade daemon built for power users, NAS hoarders, and sysadmins. It autonomously crawls massive network-attached media libraries, detects videos missing English subtitles, performs zero-disk audio extraction directly into RAM, and infers broadcast-quality .srt files using state-of-the-art faster-whisper AI models.

🛠️ Key Architectural Features

• Zero-Disk Audio Extraction: Stops SSD wear-and-tear by bypassing /tmp/ files completely. Audio streams are asynchronously ripped via FFmpeg and piped directly into RAM (NumPy arrays) for AI ingestion. Supports .mkv, .mp4, .avi, .webm, as well as broadcast and physical media formats like .ts, .m2ts, and .vob.
• Bounded Asynchronous Concurrency: Eliminates Python GIL starvation via an asyncio.TaskGroup producer-consumer pipeline. It extracts audio streams exactly as fast as the GPU can transcribe them, strictly capping memory usage.
• Intelligent Hardware Routing Matrix: Auto-detects NVIDIA GPUs (VRAM), Apple Silicon, or pure CPU environments to intelligently route to the most optimal large-v3 (Ultra tier, >= 10GB VRAM), large-v3-turbo, small, or quantized int8 model.
• NAS-Safe & Deduplicating: Backed by an asynchronous local SQLite database (WAL mode) tracking inode and size. It avoids parsing active downloads, seamlessly skips duplicate hardlinks, and performs strict POSIX Atomic os.replace operations with original MKV metadata inheritance.
• Broadcast Formatting & Progress UI: Implements a strict chunking algorithm on top of Whisper's word-level timestamps. Subtitles naturally break on terminal punctuation. The CLI features a beautiful Rich-based dynamic progress bar and a comprehensive end-of-run statistical dashboard.

---

🚀 Installation & Deployment

🐳 Docker (Recommended for TrueNAS / Unraid)

For maximum stability and ease-of-use with NVIDIA hardware, use the provided Docker stack.

1. Clone the repository:

   git clone https://github.com/arvarik/aisrt.git
   cd aisrt
   

2. Review and modify the docker-compose.yml to point to your media directory:

   volumes:
     - /mnt/user/media/movies:/media:rw
     - ./aisrt_data:/root/.config/aisrt:rw
   

3. Deploy:

   docker compose up -d
   

💻 Native Python (Ubuntu Desktop / Server / macOS)

Prerequisites: Python 3.11+ and ffmpeg must be installed on your system.

# 1. Clone repository
git clone https://github.com/arvarik/aisrt.git
cd aisrt

# 2. Create virtual environment
python3 -m venv .venv
source .venv/bin/activate

# 3. Install the application
pip install -e .

# 4. (Optional) Install development dependencies
pip install -e ".[dev]"

---

🎮 Usage

The application features a beautifully formatted CLI built on Typer and Rich.

5 Robust Usage Examples

Here are five powerful ways to utilize aisrt in your homelab or production environment:

1. The Safe Dry-Run (Check What Will Happen)

If you just pointed the tool at a massive 10TB media library and want to ensure it skips the right files without actually running the GPU inference:

aisrt scan /mnt/movies --min-age-mins 60

• What it does: Scans the library, skips files downloaded in the last 60 minutes, ignores videos that already have a .srt or an embedded English text track, and prints a beautiful table showing exactly which files will be processed.

2. Standard Mass Generation

The standard run command. Perfect for a weekly cronjob.

aisrt run /mnt/movies

• What it does: Runs the full pipeline. Extracts audio to RAM, automatically profiles your hardware, boots the optimal Whisper model (e.g., large-v3-turbo), generates broadcast-standard atomic .srt files, and saves progress to the SQLite database.

3. Foreign Cinema AI Translation (Dub to English)

You have a folder full of Anime or French cinema and want English subtitles generated.

aisrt run /mnt/anime --translate

• What it does: Same as the standard run, but it passes task="translate" into Whisper. The AI will ingest the foreign audio and natively dub it into perfect English .srt files!

4. The 24/7 NAS Watchdog Daemon

You want aisrt to run in the background on your server and automatically process new movies as soon as they are added to your library.

aisrt run /mnt/media --watch --watch-interval 30

• What it does: Turns the CLI into a long-running daemon. It will scan the directory, process any missing subtitles, and then gracefully sleep for 30 minutes before waking up to check for new files again.

5. Hardcoding Low-End Hardware

You are running this on a potato CPU or a server with a very weak GPU, and you want to ensure it doesn't try to load a heavy model, or you want to force CPU execution.

aisrt run /mnt/media --force-device cpu --force-model small.en

• What it does: Bypasses the auto-profiler and explicitly locks the execution to your CPU using the lightweight small.en quantized model.

Docker Environment Variables

If you are deploying via Docker Compose, you can configure these exact same behaviors using pydantic-settings environment variables instead of CLI flags:

• HF_TOKEN=your_huggingface_token (Bypass HuggingFace rate limits for model downloads)
• AISRT_TRANSLATE=True (Auto-dub foreign audio into English)
• AISRT_WATCH=True (Run 24/7 as a daemon)
• AISRT_WATCH_INTERVAL_MINS=60 (Time between library scans)
• AISRT_FILTERS__MIN_AGE_MINS=30 (Skip files currently being downloaded)
• AISRT_HARDWARE__FORCE_MODEL=large-v3-turbo

---

🏗️ Open Source Development

We welcome contributions! The codebase strictly adheres to enterprise-level typing and styling.

Development Setup:

poetry install   # Or pip install -e ".[dev]"

Running Tests & Linters:

ruff check .           # Linter
ruff format .          # Formatter
mypy src/aisrt tests  # Strict Type Checking
pytest tests           # Asynchronous Unit Tests

---

📜 License

Distributed under the MIT License. See LICENSE for more information.