{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:3OXKAQ6VYA4BGR6U64NAECXBJT","short_pith_number":"pith:3OXKAQ6V","schema_version":"1.0","canonical_sha256":"dbaea043d5c0381347d4f71a020ae14cf72656b722cb1666668063eadfa84f06","source":{"kind":"arxiv","id":"2605.13087","version":1},"attestation_state":"computed","paper":{"title":"Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Reverse multi-stage fine-tuning lets a 244M Whisper model match or exceed 769M counterparts on a tiered Indic speech benchmark.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aditya Srinivas Menon, Arghya Bhattacharya, Kavya Manohar, Kumarmanas Nethil, Kush Juvekar","submitted_at":"2026-05-13T06:55:55Z","abstract_excerpt":"Fine-tuning multilingual ASR models like Whisper for low-resource languages often improves read speech but degrades spontaneous audio performance, a phenomenon we term studio-bias. To diagnose this mismatch, we introduce Vividh-ASR, a complexity-stratified benchmark for Hindi and Malayalam across four tiers: studio, broadcast, spontaneous, and synthetic noise. Through a controlled study of learning-rate timing and curriculum ordering, we find that early large parameter updates improve global WER by 12 absolute points, while a hard-to-easy curriculum adds gains for spontaneous speech. These fin"},"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":true},"canonical_record":{"source":{"id":"2605.13087","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-sa/4.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T06:55:55Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"4409553652e432a0288fbbc3685db26c0fe22ef9e9445256fb4cce3bad0e9a8e","abstract_canon_sha256":"42d96ca1991276f52c29b0feee8ceb1d563f5b451840ccbff07e65fa8b32a2e0"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T03:08:58.522738Z","signature_b64":"PFGfu4g6Uls8A/VsgM3ngl78Ey7hJ7j21+xFTKDhyVkYZs9DbIRl+qmVM83zp0lIPpGZY+tyY+hsmS6OFvlCBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"dbaea043d5c0381347d4f71a020ae14cf72656b722cb1666668063eadfa84f06","last_reissued_at":"2026-05-18T03:08:58.522237Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T03:08:58.522237Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Reverse multi-stage fine-tuning lets a 244M Whisper model match or exceed 769M counterparts on a tiered Indic speech benchmark.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Aditya Srinivas Menon, Arghya Bhattacharya, Kavya Manohar, Kumarmanas Nethil, Kush Juvekar","submitted_at":"2026-05-13T06:55:55Z","abstract_excerpt":"Fine-tuning multilingual ASR models like Whisper for low-resource languages often improves read speech but degrades spontaneous audio performance, a phenomenon we term studio-bias. To diagnose this mismatch, we introduce Vividh-ASR, a complexity-stratified benchmark for Hindi and Malayalam across four tiers: studio, broadcast, spontaneous, and synthetic noise. Through a controlled study of learning-rate timing and curriculum ordering, we find that early large parameter updates improve global WER by 12 absolute points, while a hard-to-easy curriculum adds gains for spontaneous speech. These fin"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"reverse multi-stage fine-tuning (R-MFT), a training recipe that enables a parameter-efficient 244M Whisper model to match or exceed conventionally fine-tuned 769M counterparts.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the four complexity tiers in Vividh-ASR sufficiently represent the distribution of real-world usage for Indic ASR and that the observed gains from early large updates and hard-to-easy ordering will hold for other languages, models, and deployment conditions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Vividh-ASR benchmark and reverse multi-stage fine-tuning enable smaller Whisper models to match larger ones on complex Indic speech by concentrating adaptation in the decoder.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reverse multi-stage fine-tuning lets a 244M Whisper model match or exceed 769M counterparts on a tiered Indic speech benchmark.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"40fedcc19d37142dc54572e97da6291af7e71c62e5bb27f253ae6ecca315026d"},"source":{"id":"2605.13087","kind":"arxiv","version":1},"verdict":{"id":"0b47ebdd-c55d-475e-a66b-9deaec130015","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:34:16.090899Z","strongest_claim":"reverse multi-stage fine-tuning (R-MFT), a training recipe that enables a parameter-efficient 244M Whisper model to match or exceed conventionally fine-tuned 769M counterparts.","one_line_summary":"Vividh-ASR benchmark and reverse multi-stage fine-tuning enable smaller Whisper models to match larger ones on complex Indic speech by concentrating adaptation in the decoder.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the four complexity tiers in Vividh-ASR sufficiently represent the distribution of real-world usage for Indic ASR and that the observed gains from early large updates and hard-to-easy ordering will hold for other languages, models, and deployment conditions.","pith_extraction_headline":"Reverse multi-stage fine-tuning lets a 244M Whisper model match or exceed 769M counterparts on a tiered Indic speech benchmark."},"references":{"count":29,"sample":[{"doi":"","year":null,"title":"However, zero-shot word er- ror rates (WER) for many Indic languages often exceed 100%","work_id":"754b1366-824f-4b59-9833-996af328cd91","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition","work_id":"88af3e95-a33c-4a1d-ac40-39f26b850a59","ref_index":2,"cited_arxiv_id":"2605.13087","is_internal_anchor":true},{"doi":"","year":null,"title":"The Vividh-ASR Benchmark Vividh-ASR is a diagnostic benchmark organized byacous- tic and prosodic complexityrather than by domain. It targets Hindi and Malayalam, representing the Indo-Aryan and Dra- ","work_id":"1d37ff9e-a459-45fa-b4cc-90fee52a2c56","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"However, when adapting to low-resource languages with complex phonotactics, the model Table 1:Data distribution in hours","work_id":"0f87862e-9f2b-4b95-ac96-40e47d1db4ce","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Learning Rate Effect Figure 2 shows training loss for the Malayalam Whisper- medium model (representative; Hindi and Whisper-small ex- hibit identical trends)","work_id":"664c92a0-e23e-469d-8f2b-717ec1daef45","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":29,"snapshot_sha256":"a341abfc06181fb2e37f36317e2b8b9bd70f4aa51171b4f6b7d8c92ea96a894b","internal_anchors":1},"formal_canon":{"evidence_count":2,"snapshot_sha256":"a5bfea85e454db7e96a09fc042f76975640248490853e796bc5000f734e5638e"},"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":"2605.13087","created_at":"2026-05-18T03:08:58.522302+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13087v1","created_at":"2026-05-18T03:08:58.522302+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13087","created_at":"2026-05-18T03:08:58.522302+00:00"},{"alias_kind":"pith_short_12","alias_value":"3OXKAQ6VYA4B","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"3OXKAQ6VYA4BGR6U","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"3OXKAQ6V","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2605.13087","citing_title":"Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition","ref_index":2,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT","json":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT.json","graph_json":"https://pith.science/api/pith-number/3OXKAQ6VYA4BGR6U64NAECXBJT/graph.json","events_json":"https://pith.science/api/pith-number/3OXKAQ6VYA4BGR6U64NAECXBJT/events.json","paper":"https://pith.science/paper/3OXKAQ6V"},"agent_actions":{"view_html":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT","download_json":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT.json","view_paper":"https://pith.science/paper/3OXKAQ6V","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13087&json=true","fetch_graph":"https://pith.science/api/pith-number/3OXKAQ6VYA4BGR6U64NAECXBJT/graph.json","fetch_events":"https://pith.science/api/pith-number/3OXKAQ6VYA4BGR6U64NAECXBJT/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT/action/storage_attestation","attest_author":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT/action/author_attestation","sign_citation":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT/action/citation_signature","submit_replication":"https://pith.science/pith/3OXKAQ6VYA4BGR6U64NAECXBJT/action/replication_record"}},"created_at":"2026-05-18T03:08:58.522302+00:00","updated_at":"2026-05-18T03:08:58.522302+00:00"}