Tim's media tools

A small collection of useful tools for media production, made during my internship at the VPRO.

On a remote server, run he webserver with

uv run --package gateway uvicorn gateway.main:app --host 0.0.0.0 --port 8000

edl-to-archive

Simple script to convert EDL and archive source lists to complete archive lists with timecodes and source links.

Exclusion Rules

Exclusion rules allow you to filter out EDL entries before processing. Rules are defined in a text file, one rule per line.

field_name OPERATOR "value"

Operators:

• IS - exact match (case-sensitive)
• INCLUDES - substring match (case-sensitive)

Logical operators:

• AND - both conditions must be true
• OR - either condition must be true
• NOT - negates the following expression
• () - parentheses for grouping

Lines are OR'd together: if ANY line matches, the entry is excluded.

Available Fields

Field Name	Aliases	Description
name	Name, NAME	Clip name
file_name	FileName, filename, Bestandsnaam	Source file name
reel	Reel, REEL	Reel identifier
track	Track, TRACK	Track name
comment	Comment, COMMENT	Entry comment

green-to-red music

Simple tool that converts a Spotify playlist into downloaded MP3 files. Give it a Spotify playlist URL and it will find each track on YouTube, download them as MP3s, and rename them to the standard "Artist - Title" format.

Also generates a detailed CSV file with licensing metadata (composers, labels, ISRCs) sourced from MusicBrainz.

Built on SysGarcia's Playlist-converter and pH-7's Download-Simply-Videos-From-YouTube.

How it works

1. Fetches all tracks from the given Spotify playlist via web scraping (no API credentials needed)
2. Searches YouTube for each track (by track name + artist) in parallel
3. Downloads every matched video as an MP3 using yt-dlp, while simultaneously looking up licensing metadata (artist formatting, composers) on MusicBrainz
4. Renames downloaded files to "Artist - Title.mp3" format
5. Outputs a CSV with full song metadata for licensing purposes

No API credentials needed — the tool uses web scraping for Spotify playlists, youtube-search for YouTube lookups, and yt-dlp for downloads.

yt-bulk-dl

Python script that batch-downloads YouTube videos in the highest available quality. Reads URLs from a text file, merges the best video + audio streams, downloads multiple videos in parallel, and optionally fetches manually-added subtitles as .srt sidecar files.

metadata.csv contains one row per URL from download-list.txt (in input order), with columns: filename, youtube_title, channel, upload_date, youtube_url. The order of rows always matches the order of URLs in download-list.txt regardless of which download finishes first.

Notes

• Auto-generated subtitles are excluded; only manually-added subs are downloaded.
• Playlist links download only the single linked video (set noplaylist: False in the script to change this).

transcribe

A minimal transcription tab. Upload a video or audio file and get back a transcript as plain text (.txt) or subtitles (.srt). The heavy lifting (running Whisper on a GPU) is done by a separate, self-hosted transcription service that I build myself (a WhisperX-based microservice in its own repo), which exposes an OpenAI-compatible audio API and runs in its own GPU LXC. Beyond basic transcription it also supports speaker diarization and word-accurate timestamps, though this minimal tab only uses plain transcription. This tab handles the upload, the audio extraction, talking to that service, and showing progress.

How it works

1. You upload a media file (video or audio).
2. The tab immediately transcodes it to a small Whisper-friendly audio file (16 kHz mono) with ffmpeg, and deletes the original upload to save space on the server (which has limited storage).
3. The extracted audio is sent to the transcription service, which exposes an OpenAI-compatible audio API (the same POST /v1/audio/transcriptions endpoint and formats as OpenAI's own Whisper API), authenticated with a shared bearer token.
4. The tab runs that call inside a background job and shows live progress in the browser, following the same job-status pattern as green-to-red and yt-bulk-dl.
5. When the job is done, you download the transcript as .srt or .txt.

The service loads models on demand and unloads them after an idle timeout to free VRAM (the GPU is shared with other services), so the first request after a quiet period is slower (the model has to load) and the service may briefly return HTTP 429 while it's busy — the tab handles both gracefully rather than failing the job.

This tab does not run Whisper itself and has no GPU dependency — it can run on the same office-grade server as the rest of the collection. The transcription service can run anywhere reachable on the network (currently a dedicated GPU LXC on a second Proxmox host; because it's OpenAI-compatible it could later be repointed at cloud Whisper or another compatible server without changing this tab).

Configuration

Variable	Purpose
TRANSCRIBE_SERVICE_URL	Base URL of the transcription service (OpenAI-style, e.g. ends in /v1)
TRANSCRIBE_API_TOKEN	Shared bearer token / API key sent with every request
TRANSCRIBE_MODEL	Whisper model id the service should use (e.g. a large-v3-turbo CTranslate2 model)

Because the service speaks the OpenAI audio API, this tab can talk to it with the official openai client (point base_url at TRANSCRIBE_SERVICE_URL) or a plain HTTP client — and it could be repointed at OpenAI's hosted Whisper or any other compatible server without code changes.

If TRANSCRIBE_SERVICE_URL is unset the tab still loads but shows a "transcription service not configured" message instead of accepting uploads.