A preprint has appeared on arXiv proposing a method called EvoSpeakerMem: treat voiceprints as observations of identity, not identity itself. It does three things—causally routes speakers to the corresponding user entities, selectively updates acoustic states, and reads and writes memories separately by identity, while allowing the underlying memory store to be swapped out later. The authors also created the LoCoMo-ID benchmark, which includes hidden speaker identities, shifts across recording conditions, and tests involving confusion among four people.
The results: on the longitudinal SMIIP-TV benchmark, with the threshold frozen near the enrollment threshold, the genuine-speaker rejection rate rose from 29.8% to 55.6% over roughly a year. On speaker-sensitive question answering in LoCoMo-ID, gated updates outperformed fixed representations by 3.86 and 11.62 percentage points, respectively, with no erroneous user authorizations. But the authors also point out that reduced fragmentation and better question answering can coexist with incorrect cross-user links; QA alone cannot reveal the fragmentation–contamination trade-off.
$BTC
The results: on the longitudinal SMIIP-TV benchmark, with the threshold frozen near the enrollment threshold, the genuine-speaker rejection rate rose from 29.8% to 55.6% over roughly a year. On speaker-sensitive question answering in LoCoMo-ID, gated updates outperformed fixed representations by 3.86 and 11.62 percentage points, respectively, with no erroneous user authorizations. But the authors also point out that reduced fragmentation and better question answering can coexist with incorrect cross-user links; QA alone cannot reveal the fragmentation–contamination trade-off.
$BTC
