{"paper":{"title":"TW-Sound580K: A Regional Audio-Text Dataset with Verification-Guided Curation for Localized Audio-Language Modeling","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A verification-curated Taiwanese audio-text dataset and dynamic arbitration strategy lifts audio-language model accuracy on localized speech from 42.6 to 49.1 percent.","cross_cats":[],"primary_cat":"cs.SD","authors_text":"Hao-Hui Xie, Ho-Lam Chung, Hung-yi Lee, Ke-Han Lu, Wenze Ren, Xie Chen, Yi-Cheng Lin","submitted_at":"2026-03-05T12:07:19Z","abstract_excerpt":"Large Audio-Language Models (LALMs) typically struggle with localized dialectal prosody due to the scarcity of specialized corpora. We present TW-Sound580K, a Taiwanese audio-text instruction dataset developed through a Verify-Generate-Critique (VGC) protocol. This pipeline leverages Dual-ASR validation to filter 522K raw clips, subsequently expanding them into 580,000 high-fidelity instruction pairs using a teacher model. The dataset's utility is demonstrated through Tai-LALM, which fine-tunes a DeSTA 2.5-Audio-initialized backbone and incorporates a dynamic Dual-ASR Arbitration strategy to o"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On the TAU Benchmark, Tai-LALM reaches 49.1% accuracy, marking a 6.5% absolute improvement over the zero-shot baseline (42.6% with ASR text conditioning). This confirms that integrating regional corpora with rigorous curation and dynamic arbitration significantly enhances LALM performance on localized speech.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the Verify-Generate-Critique protocol combined with Dual-ASR validation produces genuinely higher-fidelity instruction pairs that causally drive the observed benchmark gain rather than other unstated differences in training or evaluation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TW-Sound580K dataset plus Tai-LALM model with dynamic Dual-ASR arbitration lifts localized Taiwanese audio-language accuracy to 49.1% on the TAU benchmark.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A verification-curated Taiwanese audio-text dataset and dynamic arbitration strategy lifts audio-language model accuracy on localized speech from 42.6 to 49.1 percent.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7ccd80476394bdec8bf5cb7d768ec5865b0577f92801ffbee3586ac984f322d4"},"source":{"id":"2603.05094","kind":"arxiv","version":3},"verdict":{"id":"5a8d4b94-47b6-43ca-9c3f-1bd09542495f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T15:36:28.639447Z","strongest_claim":"On the TAU Benchmark, Tai-LALM reaches 49.1% accuracy, marking a 6.5% absolute improvement over the zero-shot baseline (42.6% with ASR text conditioning). This confirms that integrating regional corpora with rigorous curation and dynamic arbitration significantly enhances LALM performance on localized speech.","one_line_summary":"TW-Sound580K dataset plus Tai-LALM model with dynamic Dual-ASR arbitration lifts localized Taiwanese audio-language accuracy to 49.1% on the TAU benchmark.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the Verify-Generate-Critique protocol combined with Dual-ASR validation produces genuinely higher-fidelity instruction pairs that causally drive the observed benchmark gain rather than other unstated differences in training or evaluation.","pith_extraction_headline":"A verification-curated Taiwanese audio-text dataset and dynamic arbitration strategy lifts audio-language model accuracy on localized speech from 42.6 to 49.1 percent."},"references":{"count":44,"sample":[{"doi":"","year":2026,"title":"TW-Sound580K: A Regional Audio-Text Dataset with Verification-Guided Curation for Localized Audio-Language Modeling","work_id":"6bb3cbea-64cd-42fe-9dce-42095e900ba7","ref_index":1,"cited_arxiv_id":"2603.05094","is_internal_anchor":true},{"doi":"","year":null,"title":"Background Foundational corpora like AudioSet [11] and LibriSpeech","work_id":"b257d624-c32d-4720-9320-0f2e20fc8f86","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"acoustic long-tail","work_id":"fd705707-c127-46ca-bfed-6a6142fe9067","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Methodology To bridge the localization gap, our data-centric pipeline is struc- tured into four key stages: (I) Dataset Construction, (II) Train- ing Data Generation, (III) the multimodal Training Pro","work_id":"9b9c31f1-05fe-4f1d-b5e4-34193e1f2072","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"To preserve speech-free soundmarks, clips where both ASRs yield empty outputs bypass the text check","work_id":"f0d7d77f-88ee-4324-b76d-6d1595bcbf34","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":44,"snapshot_sha256":"812665e115ada450767d90a34e3ccb747a8b2d073e2b60e42baa072addb7034c","internal_anchors":6},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ce8cde31f26c05a060d8c8f3e25dbaf7445b98f2eedbd1df5f91db579f42d99d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}