Compressing 300+ hours of a D&D series into a retelling for actors
Сбер
Anthropic
An enthusiast used AI to process 55 episodes of a Russian-language D&D actual play series, totaling over 300 hours. He built a knowledge base and wrote a video retelling for actors. The pipeline involved transcription with GigaAM v3, script generation with Claude Opus 5, and manual correction.
An AI enthusiast processed 55 episodes of the Russian-language D&D series 'Gryadut priklyucheniya' with a total duration of 302.6 hours. He transcribed the audio using GigaAM v3 from Sber, which outperforms Whisper large-v3 on Russian benchmarks, and then used Claude Opus 5 to generate structured scripts from the transcripts. He encountered problems with named entity recognition in fantasy settings, leading to incorrect names like 'Volgograd' for 'Dolgograd'. He built a glossary of 838 canonical names and corrected errors manually. The final knowledge base included 391 files with over 1.8 million words. The process took about eight days.
- Abbreviations
- D&D = Dungeons & Dragons — Dungeons & Dragons (настольная ролевая игра)
- WER = Word Error Rate — –
- LLM = Large Language Model — –
Source: Habr — хаб ML —
original