Выпуск

Updating Speech Models Weekly Without Forgetting

My Weird Prompts · 17 ноября 2026 г. · 24 мин

Frequent, lightweight updates to a speech recognition model require separating list problems from weights problems. Runtime biasing frameworks like BR-ASR handle up to 200,000 custom vocabulary entries without touching model weights. When accents or systematic errors do require training, AWS AI Labs' sequential-editing recipe uses ten epochs at one-tenth the learning rate. To prevent catastrophic forgetting across dozens of cycles, researchers use adapters, wake-and-sleep consolidation, TIES-merging, and SVD-rehearsal requiring just one utterance per past task. Key sources: • Sequential Editing for Lifelong Training of Speech Recognition Models, AWS AI Labs (Kulshreshtha et al.), arXiv v1, submitted 2024-06; retrieved 2026-11-17. — https://arxiv.org/abs/2406.17935v1 • Continuously Learning New Words in Automatic Speech Recognition, KIT / CMU (Ugan, Pham, Waibel et al.), arXiv v2; retrieved 2026-11-17. — https://arxiv.org/html/2401.04482v2 • Improving a personal ASR model by using it every day, repo README, last updated 2026-11-04; retrieved 2026-11-17. — https://github.com/danielrosehill/ASR-Recursive-Improvement-Loop Episode 6659 — myweirdprompts.com/n/6659 (permanent short link: myweirdprompts.com/078468) Listen on Android: MWP FM Player — myweirdprompts.com/apk

Страница выпуска