New speaker diarization method tightens boundaries by 70% without clean labels
Causal-anticausal co-training recovers 70% of ideal tightening from loose training data
Speaker diarization models typically train on multi-talker conversational ASR data, which prioritizes semantic continuity over precise segment boundaries. This results in loose labels—speech segments include pauses and margins that hurt downstream tasks. A new paper from Horiguchi et al., accepted to Interspeech 2026, introduces a co-training approach that enforces tight predictions without requiring manually cleaned labels.
The key insight: causal and anticausal models (one processing forward in time, the other backward) are inherently unable to reproduce the loose boundaries learned from standard training data. By using these models to generate pseudo labels, and then iteratively refining both models and labels in a co-training loop, the system progressively tightens boundaries. Experiments show the method recovers ~70% of the improvement achievable with ideal tight labels, leading to better downstream ASR and diarization accuracy. This technique is especially valuable for real-time applications where sharp segment boundaries are critical.
- Uses causal and anticausal models that cannot learn loose boundary behavior from training data
- Co-training scheme iteratively tightens pseudo labels and updates both models for progressive refinement
- Recovers 70% of tightening effect compared to ideal tight-label training, improving downstream performance
Why It Matters
Enables tighter speaker segments without manually labeled boundaries, improving accuracy for real-time diarization and ASR systems.