Determine whether Ruby decoding relies on speech evidence

Determine whether the autoregressive Ruby-ASR decoder uses speech evidence beyond preceding orthographic context when predicting lexical readings from speech-conditioned ruby-form sequences.

Background

Ruby-ASR predicts a span-bound ruby sequence in which orthographic spans are followed by their realized lexical readings. Because each reading follows its corresponding orthographic material, the autoregressive decoder may exploit preceding orthographic context rather than relying on acoustic evidence. The paper reports improved reading recognition but does not resolve whether the decoder causally uses speech information beyond that orthographic context.

Establishing this would require a controlled intervention that separates speech-conditioned evidence from information available in the preceding orthographic sequence, such as an ablation or counterfactual evaluation. The paper explicitly leaves this causal attribution unresolved and identifies it as outside the scope of the reported experiments.

References

These results do not establish whether the AR decoder uses speech evidence beyond preceding orthographic context; answering that question requires a controlled intervention outside this paper's scope.

— Ruby-ASR: Evidence-Preserving Supervision for Joint Orthographic and Lexical-Reading Recognition  (2609.27289 - Shi et al., 23 Sep 2026) in Section 5.2, “Lexical-reading recognition”