Text only partially determines prosodic realization, leaving residual variation that may align with speaker-level metadata such as regional accent or first-language (L1) background. However, because these labels are assigned to speakers, their predictive gains may overlap with broader speaker-associated structure rather than uniquely reflect category-level prosodic effects. We frame this as an attribution problem and propose a speaker-controlled audit using matched text-residual prosody predictors that differ only in access to metadata labels and speaker identity. The audit measures label utility beyond text, speaker utility beyond text, and the additional empirical utility of the label when speaker conditioning is already available. Across VCTK regional-accent labels and L2-ARCTIC L1 labels, metadata yields modest gains over text, while speaker conditioning yields larger gains and leaves little additional label utility under speaker control. Random speaker-group and label-perturbation controls further characterize the specificity and use of the supplied metadata. Importantly, speaker-held-out evaluation shows that metadata transfer to unseen speakers is limited and target-dependent, indicating that observed-speaker metadata gains do not by themselves establish speaker-independent category-level effects. These results position speaker-controlled auditing as a prerequisite attribution check before assigning category-level interpretations to metadata-conditioned predictive gains.

Replay-free continual automatic speech recognition (ASR) requires adapting to new speech domains while maintaining a single persistent model without old data, task identifiers, or inference-time routing. We identify commit-stage update admission as a distinct post-training problem for continual low-rank adaptation (LoRA): after a task-wise LoRA update has been trained, the system must decide which part of that update should become persistent in the deployed model. This decision is especially important in ASR, where domains share phonetic and transcription structure but still differ in speakers, channels, noise, speaking style, and lexical coverage. We propose HarmoLoRA, a rank-constrained commit operator that decomposes the trained current-task update with respect to previously committed directions, constructs a rank-closed residual commit path, and uses a lexicographic constrained rule to choose the smallest residual admission coefficient among near-best realized candidates. The selected update is committed to one merged ASR checkpoint, requiring no replay, task identifiers, routing, gating, or multi-branch selection at inference. Across the reported ASR orders, HarmoLoRA improves recognition performance and reduces forgetting relative to reproduced replay-free and LoRA-based continual adaptation baselines.

Contrastive decoding (CD) provides an effective inference-time strategy for audio-visual speech recognition (AVSR) by contrasting audio-visual and audio-only predictive distributions without retraining the underlying model. Existing approaches apply a fixed contrastive term at every decoding step, despite substantial variation in how informative the contrastive signal is across local decoding contexts. In AVSR, a fixed adjustment that benefits some decoding steps may provide little value or introduce instability in others, making uniform CD sensitive to the contrastive coefficient 𝛼 and limiting its reliability across varying acoustic conditions. We propose Gated CD (GCD), a training-free framework that adapts the contrastive term according to local candidate geometry at each decoding step. Since contrastive adjustment primarily affects ranking among nearby competing tokens, GCD estimates local reliability through two complementary quantities — a margin support term and a dispersion risk term — that together capture whether the contrastive signal is directionally consistent and locally stable. These are combined into a continuous gate that adaptively modulates the contrastive term during decoding. Experiments on LRS3 and LRS2 across multiple noise conditions and model scales demonstrate that GCD consistently improves over both standard AVSR decoding and uniform CD, while substantially reducing sensitivity to 𝛼, improving robustness across acoustic conditions.