Выпуск

Context-Driven LoRA Mixing for Speech Recognition

My Weird Prompts · 19 ноября 2026 г. · 23 мин

No published speech recognition system lets an AI agent dial LoRA adapters—like eighty percent tech jargon and twenty percent code-switching—from app context, and those percentages aren't real weights anyway. The discussion unpacks why NeMo's adapter API only switches or chains bottleneck adapters, why encoder LoRA hurt Parakeet-TDT in a dysarthric-speech study, and how confidence arbitration in DAS, decode-time biasing, and LLM tools like PEFT, TIES, and X-LoRA actually handle multi-adapter composition. Episode 6682 — myweirdprompts.com/n/6682 (permanent short link: myweirdprompts.com/969197) Listen on Android: MWP FM Player — myweirdprompts.com/apk

Страница выпуска