Applications of AI in news and media
summary
BBC converts 100 years of HD-equivalent material into searchable text and tags, running an in-house transcription pipeline because third-party costs are unsustainable for a 23 PB archive. Legacy Kaldi yields about 25% WER; OpenAI’s Whisper medium ~11.5% WER cuts errors near human quick-transcription levels but is 4× real time, lacks diarization and GPU support. A restoration pipeline denoises, removes physical damage, upsamples and colorizes using encoder‑decoder and attention models fine‑tuned on 60k human‑restored images, and a GPT-J fine-tuned on 500k articles provides semantic search and metadata generation.
AI-written summary · about
chapters (9)
- 0:00AI at BBC
- 1:35AI ubiquity and archive challenge
- 3:24Transcribing BBC archive
- 5:29Whisper speech-to-text
- 8:24NLP for tagging and search
- 9:54AI video restoration
- 13:13Audio enhancement plans
- 15:18GPT BBC for metadata
- 16:56Q&A
related talks
29mAI in Media: Balancing Innovation, Quality, and Responsibility
28mAI’s Next Great Challenge: How to create AI that truly understands us?
22mThe media processing pipelines behind AI
26mThe Revolutionary Work of AI & Machine Learning in Media
46mArtificial intelligence to the rescue of accessibility
16mThe End of Awkward AI Transcriptions
17mAI/ML demo derby
22mDeep learning module recognizes video sounds