The day in conversation for 2026-11-03

discussion

Transcript

The discussion: Rosa: That DM-ASR work is compelling because by conditioning dialogue generation on those diarization cues, you aren't just recognizing speech; you’re modeling the intent and speaker context simultaneously. How much of that temporal conditioning do you think is actually learned versus what’s being explicitly fed in? Dev: I find the reliance on those cues interesting, especially for multi-speaker setups. But the real question is scalability; if we try to push those compact models onto edge devices, how does that dependency on accurate diarization cues affect inference latency compared to a purely acoustic model? We saw earlier that even small increases in model complexity can derail deployment timelines. Taro: Speaking of complexity, I was looking at the segmentation papers—the RC-aware nnU-Netv2 achieving second place in BraTS 2025 is impressive given the cavity awareness. But does that high level of spatial detail ever conflict with the noise constraints we discussed last week regarding millimeter-wave transmission? For instance, if we try to apply that high-

More episodes

← Home