Datasets:

Darayut
/

sota-eval-data

license: cc-by-sa-4.0
task_categories:
  - automatic-speech-recognition
language:
  - km
  - en
configs:
  - config_name: ddd_dev
    data_files: data/ddd_dev.parquet
  - config_name: ddd_test
    data_files: data/ddd_test.parquet
  - config_name: fleurs_km_dev
    data_files: data/fleurs_km_dev.parquet
  - config_name: fleurs_km_test
    data_files: data/fleurs_km_test.parquet
  - config_name: fleurs_en_dev
    data_files: data/fleurs_en_dev.parquet
  - config_name: fleurs_en_test
    data_files: data/fleurs_en_test.parquet
  - config_name: libri_dev-clean
    data_files: data/libri_dev-clean.parquet
  - config_name: libri_test-clean
    data_files: data/libri_test-clean.parquet
  - config_name: libri_test-other
    data_files: data/libri_test-other.parquet

SOTA evaluation sets

The nine sets that training/scripts/03_evaluate.py scores in the SOTA repo, one parquet file each. Audio is stored as the original encoded bytes (DDD and FLEURS: wav, LibriSpeech: flac); text is the reference exactly as scored.

Set	Source	Utterances
`ddd_dev`	DDD, Khmer	400
`ddd_test`	DDD, Khmer	1993
`fleurs_km_dev`	FLEURS, Khmer	326
`fleurs_km_test`	FLEURS, Khmer	771
`fleurs_en_dev`	FLEURS, English	394
`fleurs_en_test`	FLEURS, English	647
`libri_dev-clean`	LibriSpeech	2703
`libri_test-clean`	LibriSpeech	2620
`libri_test-other`	LibriSpeech	2939

Sources and licences: DDD from Digital-Divide-Data/khmer-speech-dataset (CC BY-SA 4.0), FLEURS (CC BY 4.0), LibriSpeech (CC BY 4.0). The DDD dev and test rows were held out of every training run. Use of the combined set is under the most restrictive licence (CC BY-SA 4.0).