tigrinya-asr-merged

A merged Tigrinya speech-recognition dataset, combining and deduplicating:

• badrex/tigrinya-speech (train pool)
• google/WaxalNLP config tir_asr (train pool)
• UBC-NLP/SimbaBench_dataset config asr_test_tir (held-out benchmark test set)

Processing

• Standardized to audio (16kHz mono) and text columns, with a source column tracking origin
• Unicode NFC-normalized transcripts, empty transcripts dropped
• Exact-duplicate transcripts removed from the train pool
• Train-pool rows whose transcript overlaps with the SimbaBench test set were dropped, to prevent test leakage
• Train pool re-split into train (90%) / validation (10%), ignoring original source splits
• SimbaBench's asr_test_tir split is kept unmodified as the test split

Rows & duration

• Train: 45178 clips (196.35 hours)
• Validation: 5020 clips (21.85 hours)
• Test: 7 clips (0.01 hours)
• Total: 50205 clips (218.22 hours)

License

Check the license of each source dataset before redistributing; this card assumes CC-BY-4.0 as the most permissive common denominator among the sources at time of writing — verify this still holds for all four before relying on it.