A speech recognizer can lose its connection to the audio without producing a fluent invented transcript. Researchers studying two separately trained recognition systems found that bypassing the final stage of their audio encoder disrupted almost every utterance, while bypassing middle stages had little effect. The resulting text was garbled or repetitive. Their study therefore identifies a stage needed to keep recognition groundingTying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.Full definition → in sound, but the intervention does not reproduce fluent fabrication itself. The authors describe this as a precondition for fluent false transcripts, not the complete origin of those that occur naturally.
Terms in this piece · Glossary
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
articleClosed-World Resolution Against Tool Hallucination in LLM AgentsLaxmipriya Ganesh Iyer
articleRefusing the Impossible: A Taxonomy and Benchmark for Code Hallucination in Large Language ModelsVishnu Asutosh Dasu, Ashish Kundu, Gang Tan
articleModular Cognitive Architecture Emerges in Large Language ModelsPengrui Han, Jacob Andreas, Evelina Fedorenko, Andrea Gregor de Varda