aintoniodev/video-to-text

v1.0.0MIT

Extract transcript text from videos (TikTok, YouTube, Instagram, X, any URL) or local video files — fully local, no API keys. Ships scripts/vxtract.py: yt-dlp download + ffmpeg audio extraction + faster-whisper transcription (Spanish/English/auto, CPU or CUDA) with .txt/.md/.srt output. Alternative to cloud transcription services.

GitHub

video-to-text

Extract transcript text from videos (TikTok, YouTube, Instagram, X, any URL) or local video files — fully local, no API keys.

Engine: yt-dlp (download) + ffmpeg (audio) + faster-whisper (STT).

Install

Agent Plugins compatible client, or copy the skill directly:

cp -r video-to-text/skills/video-to-text ~/.pi/agent/skills/video-to-text

Dependencies (one-time, system or venv Python):

pip install yt-dlp faster-whisper "curl_cffi<0.16"

ffmpeg on PATH (or pass --ffmpeg <path>).

Usage

python skills/video-to-text/scripts/vxtract.py https://vm.tiktok.com/XXXX --lang es --srt
python skills/video-to-text/scripts/vxtract.py https://youtu.be/XXXX --lang auto
python skills/video-to-text/scripts/vxtract.py clip.mp4 --lang es

Output: transcript to stdout + <title>.txt (or .md with --md, plus .srt with --srt).

Why faster-whisper and not canary/transcribe.cpp

Measured 2026-08-13: canary-180m-flash and canary-1b-flash (transcribe.cpp) silently return EMPTY output for Spanish on many inputs (content-dependent EOS collapse — same file: 5s ok, 6s empty, 8s ok, 9–11s empty). faster-whisper small transcribed every test case correctly, es and en, CPU and CUDA.