{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:2FHILI267UZDXHDMH4EXHKEU62","short_pith_number":"pith:2FHILI26","schema_version":"1.0","canonical_sha256":"d14e85a35efd323b9c6c3f0973a894f697f55e54454072e7307450bb8f112b7c","source":{"kind":"arxiv","id":"2605.02223","version":1},"attestation_state":"computed","paper":{"title":"Toward Fine-Grained Speech Inpainting Forensics:A Dataset, Method, and Metric for Multi-Region Tampering Localization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Partial speech inpainting at word granularity evades existing deepfake detectors, but a new iterative method recovers the tampered regions.","cross_cats":["cs.CV"],"primary_cat":"cs.SD","authors_text":"Cong Tran, Cuong Pham, Hai Nguyen, Tung Vu, Yen Nguyen","submitted_at":"2026-05-04T04:54:29Z","abstract_excerpt":"Recent advances in voice cloning and text-to-speech synthesis have made partial speech manipulation - where an adversary replaces a few words within an utterance to alter its meaning while preserving the speaker's identity - an increasingly realistic threat. Existing audio deepfake detection benchmarks focus on utterance-level binary classification or single-region tampering, leaving a critical gap in detecting and localizing multiple inpainted segments whose count is unknown a priori. We address this gap with three contributions. First, we introduce MIST (Multiregion Inpainting Speech Tamperi"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.02223","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.SD","submitted_at":"2026-05-04T04:54:29Z","cross_cats_sorted":["cs.CV"],"title_canon_sha256":"f26301377312b6624a01700500d96cc01136b28875ec794f41b5ca3b4f04b704","abstract_canon_sha256":"b8dcb68be4e3e27756e17e96db747026ee29c853c6ed33dcdbd962242dc96fbe"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T01:05:15.241678Z","signature_b64":"K5YC5t1l3NjGO8f6Iihs/KK28EJQeLkOg/L7/yNr7ykbUILMEFXwjw7h4/pzwlQzk5F6S8AFkPyWeAdzPW3dBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"d14e85a35efd323b9c6c3f0973a894f697f55e54454072e7307450bb8f112b7c","last_reissued_at":"2026-05-20T01:05:15.238799Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T01:05:15.238799Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Toward Fine-Grained Speech Inpainting Forensics:A Dataset, Method, and Metric for Multi-Region Tampering Localization","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Partial speech inpainting at word granularity evades existing deepfake detectors, but a new iterative method recovers the tampered regions.","cross_cats":["cs.CV"],"primary_cat":"cs.SD","authors_text":"Cong Tran, Cuong Pham, Hai Nguyen, Tung Vu, Yen Nguyen","submitted_at":"2026-05-04T04:54:29Z","abstract_excerpt":"Recent advances in voice cloning and text-to-speech synthesis have made partial speech manipulation - where an adversary replaces a few words within an utterance to alter its meaning while preserving the speaker's identity - an increasingly realistic threat. Existing audio deepfake detection benchmarks focus on utterance-level binary classification or single-region tampering, leaving a critical gap in detecting and localizing multiple inpainted segments whose count is unknown a priori. We address this gap with three contributions. First, we introduce MIST (Multiregion Inpainting Speech Tamperi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Zero-shot evaluation reveals that partial inpainting at word granularity remains unsolved by existing deepfake detectors: utterance-level classifiers trained on fully synthesized speech assign near zero fake probability to MIST utterances where only 2-7% of content is manipulated. ISA consistently outperforms non-iterative baselines in this challenging setting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The generated MIST utterances with LLM-guided semantic replacement and neural voice cloning accurately represent realistic adversarial partial tampering, and the gap-tolerant region proposal plus boundary refinement in ISA can recover all regions without prior knowledge of their number.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"A new dataset, iterative coarse-to-fine localization framework, and segment-level IoU F1 metric tackle the open problem of detecting multiple unknown word-level inpainted regions in speech.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Partial speech inpainting at word granularity evades existing deepfake detectors, but a new iterative method recovers the tampered regions.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b403362080e884d7e39d02667ac2b665d0b1fb57842a1a7ac66e2620ec5eb201"},"source":{"id":"2605.02223","kind":"arxiv","version":1},"verdict":{"id":"771e4193-19df-41da-84a6-1727ca5df757","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T03:03:36.029138Z","strongest_claim":"Zero-shot evaluation reveals that partial inpainting at word granularity remains unsolved by existing deepfake detectors: utterance-level classifiers trained on fully synthesized speech assign near zero fake probability to MIST utterances where only 2-7% of content is manipulated. ISA consistently outperforms non-iterative baselines in this challenging setting.","one_line_summary":"A new dataset, iterative coarse-to-fine localization framework, and segment-level IoU F1 metric tackle the open problem of detecting multiple unknown word-level inpainted regions in speech.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The generated MIST utterances with LLM-guided semantic replacement and neural voice cloning accurately represent realistic adversarial partial tampering, and the gap-tolerant region proposal plus boundary refinement in ISA can recover all regions without prior knowledge of their number.","pith_extraction_headline":"Partial speech inpainting at word granularity evades existing deepfake detectors, but a new iterative method recovers the tampered regions."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.02223/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T16:32:21.953417Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"8352c6e70af4bc0fe86af98e388b2478c92e067c68dc15a3610f24008121fc2d"},"references":{"count":12,"sample":[{"doi":"","year":null,"title":"A V- Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset.arXiv preprint arXiv:2311.15308","work_id":"b7634f93-27dc-4d33-950b-75eb0d883418","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"CosyVoice: A Scalable Multilingual Zero-shot Text-to-speech Synthesizer based on Supervised Semantic Tokens","work_id":"e5ad925a-4045-49b5-b301-208bcbf3eca8","ref_index":2,"cited_arxiv_id":"2407.05407","is_internal_anchor":true},{"doi":"","year":null,"title":"CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training","work_id":"ce66e767-526a-419d-bb95-ac019edf4050","ref_index":3,"cited_arxiv_id":"2505.17589","is_internal_anchor":true},{"doi":"","year":null,"title":"LlamaPartialSpoof: An LLM-driven fake speech dataset simulating disinformation generation.arXiv preprint arXiv:2409.14743","work_id":"4af0d88f-1862-471e-ac1a-13aabfecc2a0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"ASVspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech","work_id":"","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":8,"snapshot_sha256":"d45f441b343914b54d610a304f3cf3255db7060db947ca15c2b5b65b3f1fef99","internal_anchors":3},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.02223","created_at":"2026-05-20T01:05:15.241080+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.02223v1","created_at":"2026-05-20T01:05:15.241080+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.02223","created_at":"2026-05-20T01:05:15.241080+00:00"},{"alias_kind":"pith_short_12","alias_value":"2FHILI267UZD","created_at":"2026-05-20T01:05:15.241080+00:00"},{"alias_kind":"pith_short_16","alias_value":"2FHILI267UZDXHDM","created_at":"2026-05-20T01:05:15.241080+00:00"},{"alias_kind":"pith_short_8","alias_value":"2FHILI26","created_at":"2026-05-20T01:05:15.241080+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62","json":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62.json","graph_json":"https://pith.science/api/pith-number/2FHILI267UZDXHDMH4EXHKEU62/graph.json","events_json":"https://pith.science/api/pith-number/2FHILI267UZDXHDMH4EXHKEU62/events.json","paper":"https://pith.science/paper/2FHILI26"},"agent_actions":{"view_html":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62","download_json":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62.json","view_paper":"https://pith.science/paper/2FHILI26","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.02223&json=true","fetch_graph":"https://pith.science/api/pith-number/2FHILI267UZDXHDMH4EXHKEU62/graph.json","fetch_events":"https://pith.science/api/pith-number/2FHILI267UZDXHDMH4EXHKEU62/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62/action/timestamp_anchor","attest_storage":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62/action/storage_attestation","attest_author":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62/action/author_attestation","sign_citation":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62/action/citation_signature","submit_replication":"https://pith.science/pith/2FHILI267UZDXHDMH4EXHKEU62/action/replication_record"}},"created_at":"2026-05-20T01:05:15.241080+00:00","updated_at":"2026-05-20T01:05:15.241080+00:00"}