Hafpaf / bornHackSubs

BornHack subtitle project

BornHack subtitles

Goal: Create subtitles for talks at BornHack

Inspiration: Some chaos events use c3subtitles to creating subtitles for talks. This serves multiple purposes: enabling folks who have hearing difficulties to access talk information and providing an option to follow talks silently in the background.

There are two main tasks for this project:

1. Transcribe talks
2. Proof reading and timing correction

The transcription will largely be done with OpenAI Whisper, an open source Automatic Speech Recognition (ASR) machine learning model. I can be done locally on a machine with minimum 16 GB GPU VRAM as well as 16 GB RAM.

The proofreading involves humans to correct context and timing errors from Whisper, as well as ensuring the overall quality of the subtitles.

Contribute

Timing correction and proofreading

• Step 1) Read and try to follow the c3subtiles styleguide when creating subtitles
• Step 2) Choose a talk and check if it has already been transcribed
• Step 3) Create an issue on Github outlining which video you are working on and progress
• Step 4) Move subrip file from raw transcription directory to draft directory
• Step 5) Do proofreading and timing.
  • Use xxx tool or similar to easier process SRT files (ToDo: Find a good an easy-to-use tool)
• Step 6) If finished, move to finished directory and mark Github Issue as resolved

Transcription

Use the script collect_talks.sh to collect all talks available for a BornHack camp or simply download a single video.

Usage:

• Find the video file RSS feed from a BornHack camp in the top right corner of the media.ccc.de front-end, see BornHack 2023.
• Use the RSS feed as argument for script

./collect_talks.sh https://media.ccc.de/c/bornhack2023/podcast/mp4-hq.xml

Setup and use the ASR subtitles tool to transcribe with the OpenAI Whisper model. The large model is preferred.

Numbers

Computational numbers for future references

BornHack 2023

It took an Nvidia RTX A4000 GPU following amount of time:

• Medium model time: 1 hour and 58 minutes (2023-11-10)
• Large model time: 2 hours and 50 minutes (2023-11-12)

Following videos where transcribed

• import-56116-eng-Chat_Control_-_the_next_months_will_be_critical_hd.mp4
• import-56117-eng-Surely_FOSS_has_no_technical_debt_Right_Right_hd.mp4
• import-56118-eng-An_introduction_to_digital_consent_Why_a_new_definition_with_new_tools_and_specifications_are_needed_hd.mp4
• import-56119-eng-LabIX_Creating_an_Internet_Exchange_in_Your_Local_Hackerspace_hd.mp4
• import-56120-eng-This_years_BornHack_badge_with_NFC_hd.mp4
• import-56121-eng-Hello_World_hd.mp4
• import-56122-eng-Open_source_chip_design_hd.mp4
• import-56123-eng-funion_A_Tor_Client_in_Elixir_hd.mp4
• import-56124-eng-Goodbye_World_hd.mp4
• import-56125-eng-State_of_the_Game_hd.mp4
• import-56140-eng-Performant_cross-platform_development_using_Flutter_hd.mp4
• import-56141-eng-R_on_OpenBSD_hd.mp4
• import-56142-eng-Sexy_SSH_Hacks_hd.mp4
• import-56143-eng-SimpleX_Chat_-_Simple_Messaging_With_Unusually_Good_Privacy_hd.mp4
• import-56144-eng-All_APIs_suck_can_we_build_one_that_sucks_less_hd.mp4
• import-56145-eng-How_to_create_better_content_videos_faster_with_OBS_hd.mp4

BornHack 2022

Transcription of all recorded BornHack 2022 talks with the large Whipser model, took 4 hours and 2 minutes (2023-11-12).

• GPU: Nvidia GTX A4000

BornHack 2021

Transcription of all recorded BornHack 2021 talks with the large Whipser model, took 3 hours and 49 minutes (2023-11-14).

• GPU: Nvidia GTX A4000