A wav2vec2.0-xlsr-53 model is fine-tuned on a large speech corpus of children aged 3 to 5 years.
language:
- ko
thumbnail: "url to a thumbnail used in social sharing"
tags:
- automatic-speech-recognition
- speech
datasets:
- AI-HUB
metrics:
- cer
model-index:
- name: xlsr-53-korean_kids_3_to_5_syl
results:
- task:
type: automatic-speech-recognition
dataset:
type: AI-HUB
name: AI-HUB
split: test
metrics:
- type: cer
value: 3.48
name: Syllable Error Rate
Dataset: children's speech aged 3 to 5 years
- the number of files) train:valid:test=256,548:12,121:12,000
- transcripts were tokenized into morphs by Mecab
EVALUATION results
- Dataset: testset from the corpus
- Performance: syllable Error Rate on the testset = 3.48%