{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:BMQNALFCOUX6AIDEBQF2KY6BUJ","short_pith_number":"pith:BMQNALFC","schema_version":"1.0","canonical_sha256":"0b20d02ca2752fe020640c0ba563c1a246edbc30202e74a3653212ba390f1884","source":{"kind":"arxiv","id":"2606.21458","version":1},"attestation_state":"computed","paper":{"title":"Post-Training Speech Enhancement Language Models with Perceptual Rewards","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Antonis Asonitis, Fr\\'ed\\'eric Berdoz, Luca A. Lanzend\\\"orfer, Roger Wattenhofer","submitted_at":"2026-06-19T14:14:20Z","abstract_excerpt":"Speech enhancement language models achieve strong results when trained on discrete audio tokens, but their optimization relies on token-level cross-entropy rather than the perceptual metrics used for evaluation. We introduce a post-training stage for autoregressive speech enhancement language models using Group Sequence Policy Optimization (GSPO) with multi-metric perceptual rewards. Our method directly optimizes non-differentiable quality metrics (DNSMOS, WER, and UTMOS) as reward signals, without learned surrogates or offline preference pairs. Applied to two autoregressive base models, UniSE"},"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":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2606.21458","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-06-19T14:14:20Z","cross_cats_sorted":[],"title_canon_sha256":"bb17795a6318614fcc20f54e6202c35d2e8b1faccd90228b958b91b68b71789b","abstract_canon_sha256":"fe7c7d497a9caac3a91b925684b15a1c99e40ed7744d933fbf72e2d8fb65fc4f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-23T01:13:11.001393Z","signature_b64":"1gK1gMNvkA+vjdMqnx4nMpQIySpQBKwP7Kt8QCFH1+T3mg9eXftjzGVoimQiWi4y6ijwQOxT/O/+yYr+fUK3Dw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0b20d02ca2752fe020640c0ba563c1a246edbc30202e74a3653212ba390f1884","last_reissued_at":"2026-06-23T01:13:11.000936Z","signature_status":"signed_v1","first_computed_at":"2026-06-23T01:13:11.000936Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Post-Training Speech Enhancement Language Models with Perceptual Rewards","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Antonis Asonitis, Fr\\'ed\\'eric Berdoz, Luca A. Lanzend\\\"orfer, Roger Wattenhofer","submitted_at":"2026-06-19T14:14:20Z","abstract_excerpt":"Speech enhancement language models achieve strong results when trained on discrete audio tokens, but their optimization relies on token-level cross-entropy rather than the perceptual metrics used for evaluation. We introduce a post-training stage for autoregressive speech enhancement language models using Group Sequence Policy Optimization (GSPO) with multi-metric perceptual rewards. Our method directly optimizes non-differentiable quality metrics (DNSMOS, WER, and UTMOS) as reward signals, without learned surrogates or offline preference pairs. Applied to two autoregressive base models, UniSE"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.21458","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.21458/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"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":"2606.21458","created_at":"2026-06-23T01:13:11.001016+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.21458v1","created_at":"2026-06-23T01:13:11.001016+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.21458","created_at":"2026-06-23T01:13:11.001016+00:00"},{"alias_kind":"pith_short_12","alias_value":"BMQNALFCOUX6","created_at":"2026-06-23T01:13:11.001016+00:00"},{"alias_kind":"pith_short_16","alias_value":"BMQNALFCOUX6AIDE","created_at":"2026-06-23T01:13:11.001016+00:00"},{"alias_kind":"pith_short_8","alias_value":"BMQNALFC","created_at":"2026-06-23T01:13:11.001016+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2606.21458","citing_title":"Post-Training Speech Enhancement Language Models with Perceptual Rewards","ref_index":1,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ","json":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ.json","graph_json":"https://pith.science/api/pith-number/BMQNALFCOUX6AIDEBQF2KY6BUJ/graph.json","events_json":"https://pith.science/api/pith-number/BMQNALFCOUX6AIDEBQF2KY6BUJ/events.json","paper":"https://pith.science/paper/BMQNALFC"},"agent_actions":{"view_html":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ","download_json":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ.json","view_paper":"https://pith.science/paper/BMQNALFC","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.21458&json=true","fetch_graph":"https://pith.science/api/pith-number/BMQNALFCOUX6AIDEBQF2KY6BUJ/graph.json","fetch_events":"https://pith.science/api/pith-number/BMQNALFCOUX6AIDEBQF2KY6BUJ/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ/action/storage_attestation","attest_author":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ/action/author_attestation","sign_citation":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ/action/citation_signature","submit_replication":"https://pith.science/pith/BMQNALFCOUX6AIDEBQF2KY6BUJ/action/replication_record"}},"created_at":"2026-06-23T01:13:11.001016+00:00","updated_at":"2026-06-23T01:13:11.001016+00:00"}