Data and Code Release for "Benchmarking Multilingual ASR for Tagalog and Cebuano Across Scoring Conventions and Language Coverage" (NLPIR 2026)

Release archive for the NLPIR 2026 paper "Benchmarking Multilingual ASR for Tagalog and Cebuano Across Scoring Conventions and Language Coverage" (paper NJ 105). It contains the unnormalized per-utterance hypotheses of all fourteen runs of Whisper large-v3, MMS-1B-all, and SeamlessM4T v2 Large on the FLEURS fil_ph and ceb_ph test splits; the run configurations, environment specification, and exact model revisions; the inference and scoring notebook with the four normalization conventions, scoring, alignment, and bootstrap code and its worked-example tests; all result tables; a re-execution of the published SeamlessM4T decoding path with and without special-token suppression; the language-identity audit outputs; the labels of two independent annotators for the 46-utterance validation sample; and the Cebuano–Tagalog word list used in the substitution analysis. FLEURS audio and reference transcriptions are not redistributed and should be obtained from google/fleurs (CC BY 4.0); every file joins to it on idx or utt_id. See README.md inside the archive for details.