Why Speaker Embeddings Fail Across Time and Noise
From My Weird Prompts · 27:00
Show by Daniel Rosehill
Speaker diarization turns voices into numerical vectors, but persistent voiceprints are far more fragile than assumed. Examining architectures like ECAPA-TDNN and spectral clustering reveals how pipelines guess headcounts and why tools cannot easily track speakers across separate sessions. Benchmark data highlights massive accuracy drops between studio microphones and messy environments, while acoustic shifts like aging, cassette degradation, and stroke demonstrate why durable lifetime voice profiles remain out of reach. Key sources: • VoxAging: Continuously Tracking Speaker Aging with a Large-Scale Longitudinal Dataset in English and Mandarin (Bao, Yang, Li, Xu; 2025-05) — https://arxiv.org/abs/2505.21445 • Stabilizing Short Duration Speaker Verification through Neural Re-scoring with Hybrid Enrollment (Ai et al.; 2026-06-15; Interspeech 2026) — https://arxiv.org/abs/2606.16115v1 • ECAPA-TDNN Embeddings for Speaker Diarization (INTERSPEECH 2021) — https://www.isca-archive.org/interspeech_2021/dawalatabad21_interspeech.pdf Episode 6432 — myweirdprompts.com/n/6432 (permanent short link: myweirdprompts.com/031604) Listen on Android: MWP FM Player — myweirdprompts.com/apk
More episodes of My Weird Prompts
More episodes
- Why the Gaussian Distribution Fails in Plain Sight
- Where to Learn LLM Training Without Generic Courses
- Why Synthetic Fine-Tuning Data Needs Tooling
- The Unsung Builders of Modern Speech AI
- Choosing Between MCP Servers, Agent Skills, and CLIs
- Turning Open-Source Triage into Subscription Credit
- Who Actually Won the Open Weight Race
- Which AI Models Actually Last
- Why Recursive File Watching Strains Linux Memory
- Securing MCP Endpoints and Webhooks Without Open Doors
podsearch is a free, independent podcast search engine. Read about podsearch, the privacy policy or contact us.