Sbobinator

Sbobinator transcribes audio and video to text in Italian, locally, using the pre-trained NVIDIA NeMo Parakeet TDT 0.6B v3 model.

Bilingual documentation

This site is also available in Italiano. Use the language selector in the top navigation bar.

What it does

Feature	Description
Transcription	Audio/video → text + SRT subtitles
Job queue	Multiple files processed one at a time; list at /jobs, detail at /jobs/{id}
Summary	Multi-provider LLM (DeepSeek, OpenAI, Qwen local, …)
Web UI	FastAPI + HTMX on port 8501 (native) or 8502 (Docker)
CLI	sbobina for automation and headless server

Design principles

1. Local processing — no audio is sent to the cloud during normal use.
2. Offline models — Parakeet (ASR) and optional Qwen GGUF in models/.
3. No overwrites — each job gets its own timestamped folder.
4. Cross-platform — native Python (development) and Docker on Linux (deployment).

Hardware requirements

Requirements depend on your use case (transcription only, API summary, local summary). See the dedicated guide:

Hardware requirements — RAM, disk, network, and GPU for each mode.

Resource	Minimum	Recommended
Python	3.12+	3.12 or 3.13
RAM (ASR only)	8 GB	16 GB
RAM (local Qwen)	16 GB	32 GB
Disk	6 GB free	10 GB
ffmpeg	Required	In PATH
NVIDIA GPU	Optional	CUDA for faster ASR

Quick start (Windows, Python)

python scripts\install_local.py
python scripts\download_model.py
start.bat

Cloud summary: configure an API key at http://localhost:8501/settings/summary

Open http://localhost:8501, upload a file, click Queue transcription.

Documentation

Section	Content
Getting started	Installation and first steps
User guide	UI, CLI, queue, summaries
Architecture	Components and data flow
Deploy	Docker and environment variables
Troubleshooting	Common errors
Reference	FAQ, DB schema, glossary, licenses

Build documentation locally

pip install -r docs/requirements.txt
mkdocs serve

Site at http://127.0.0.1:8000 — publish: scripts/publish_docs.bat (or python scripts/publish_docs.py), then git push in the sibling repo sbobinator.github.io.

License

Proprietary software — Copyright © 2024-2026 Antonio Trento. Free for personal use; commercial/enterprise use requires a paid license. See LICENSE and Licenses.

NVIDIA and Hugging Face models have their own licenses.