Identifying who says what in a long TV drama is harder than it sounds — and a new paper argues that reasoning models can do it better than anyone has so far.
Researchers released DramaSR-532K, a benchmark of 532,000 annotated dialogue lines spanning more than 900 characters drawn from TV dramas. Alongside it, they propose DramaSR-LRM, a speaker-recognition system built on a large reasoning model. The system pulls together audio, text, and visual cues using multimodal tool-use to figure out who is speaking, even when the audio alone is not enough to go on. Both the dataset and code are being made publicly available.
The hard part of this problem is short utterances — a single line of dialogue, maybe two words — where acoustic fingerprinting largely fails. That is exactly where DramaSR-LRM claims its biggest gains over existing baselines. If the results hold up, it is a meaningful step toward machines that can actually follow a story rather than just transcribe one.
Speaker diarization in clean, single-speaker audio is a solved problem. Long-form drama, with overlapping characters, ambient noise, and years of storyline context, is not — and most video AI benchmarks have not even tried to measure it at this scale.