A study of automatic speech recognition has identified the final encoder stage as a critical point where audio representations become usable for grounded transcription. Researchers tested two independently trained Conformer-Large systems—one using CTC decoding and the other RNN-T—under environmental degradation and changes in speaker background.
Bypassing middle encoder blocks had little effect, but bypassing the final block caused output to diverge on nearly every utterance in both systems. At that stage, internal representations became more compact, grapheme information became explicit and the trained decoder could begin reading text from them. The repeated result across two decoder families suggests a shared terminal-stage dependency.
The intervention did not recreate natural hallucinations. Instead of fluent but unrelated sentences, it produced garbled or repetitive output. The researchers therefore describe the finding as a mechanistic precondition for hallucination: loss of adequate grounding becomes possible at this boundary, but other processes would be needed to turn that failure into plausible fabricated text.
The paper is a new preprint and examines two particular architectures, so the mechanism should not be generalized to every speech model. It nevertheless gives developers a more precise place to probe when noise or distribution shifts cause transcripts to detach from their audio.