MIT Video Productions Transcription and Summarization Pipeline

MIT Video Productions is the on-campus video production unit that video records faculty and department talks and symposia, Institute events and courses among a wide range of activities. Over its lifetime the unit has recorded over 45,000 tapes. Recently 17,000 tapes were selected for digitization (with the remaining tapes having been previously digitized or were initially recorded digitally). However, as a video production unit the metadata captured for each recording was limited to the date of recording, an event title, sometimes additional speaker or event information and some of the technical details. Most importantly, the contents of the recordings were not available for discovery.

With recent advances in machine learning and generative AI, we hypothesized that we could use contemporary tools to transcribe the video at a “good enough” level of accuracy to enable search within video contents and to automatically summarize the transcripts staff and archivists to use for discovery.

For the transcription and summarization project I developed transcription and summarizations processing pipelines running locally using macOS and OpenAI Whisper transcription and open LLMs for summarization. I utilized industry standard Word Error Rate metrics on representative video to conclude we could expect “accurate enough” transcripts for internal search and discovery purposes. I then processed over 10,000 videos to create text transcripts, caption files and automatic content summaries.

Info

Massachusetts Institute of Technology
MIT Open Learning
2024-2026

Key Collaborators: George Blood (George Blood, LP), Laura Brown* (George Blood LP), Clayton Hainsworth* (MIT, Video Productions), Brigette Tersek* (MIT, Video Productions)

Funding Source: MIT Open Learning