Audio deepfake detection is commonly formulated as clip-level binary classification of single-domain audio. However, real-world manipulated audio can exhibit mixed authenticity, where genuine and manipulated cues coexist across temporal transitions, overlapping sources, or both. This setting requires not only detecting manipulated audio but also localizing the components that provide evidence for the decision. We propose ToolDF, a tool-integrated reasoning framework for mixed-authenticity audio deepfake detection. ToolDF employs an audio large language model as an orchestrator trained with supervised tool-use trajectories. It adaptively analyzes the audio scene, selectively performs source separation, routes components to domain-specific experts, and aggregates their evidence into an interpretable verdict. We further introduce a mixed-authenticity ADD benchmark covering temporal transitions, acoustic overlaps, and hybrid mixtures. Experimental results show that ToolDF achieves the best overall performance on composite-type detection, achieving macro-F1 gains of 3.72 and 14.39 points over the strongest monolithic baseline and a fixed pipeline, respectively, while providing interpretable evidence localized to temporal regions and acoustic sources. Our source code and dataset are publicly available online at https://github.com/rlataewoo/tooldf.

Automated pronunciation assessment predicts pronunciation quality at the phone, word, and utterance levels. Recent joint models use both canonical and realized phone representations, but their mismatch is often encoded implicitly rather than preserved as a separate scoring path. This makes it difficult to trace how local phone errors affect word- and utterance-level scores. We propose HiGraph, which explicitly computes canonical-realized mismatch and propagates it through a hierarchical phone-to-word-to-utterance graph. This creates a structured discrepancy path from phone-level mismatch evidence to higher-level score representations. Under the same pipeline as our reproduced HMamba baseline, HiGraph improves word stress and utterance completeness on SpeechOcean762 (SO762). The learned word-level trace score also aligns with mispronunciation detection and diagnosis (MDD) ground truth. When words are ranked by this trace score, the top-ranked word contains a true mispronunciation in 81.5% of test utterances, and at least one of the top two words does so in 93.8% of test utterances. These results show that explicit discrepancy propagation can improve selected assessment dimensions while supporting feedback localization grounded in annotated mispronunciations.