Interspeech 2022 · First author · Sep 2022
K-Wav2vec 2.0: Korean ASR with Grapheme–Syllable Joint Decoding
01Background & Goals
- The paper explores and optimizes the components of Wav2vec 2.0 for Korean and proposes K-Wav2vec 2.0.
- Korean syllables are composed of graphemes, so using both units together can improve accuracy and vocabulary coverage at once.
02Contributions
Grapheme–syllable multi-task head and joint decoder
- Designed a multi-task architecture that learns grapheme and syllable outputs on top of pre-trained Wav2vec 2.0
- Combined the beam-search scores of both outputs in joint decoding to reflect the syllable–grapheme relation
Further pre-training and experiments
- Further pre-trained an English-pretrained model on unlabeled Korean speech
- Built the experiment setup in Fairseq and ESPnet and evaluated on KsponSpeech and ClovaCall
03Tech Stack
- Framework / Platform
- PyTorch, Fairseq, ESPnet
- Methodology
- Self-supervised learning, Multi-task learning, Joint decoding
04Results
- Published at Interspeech 2022 (first author); code released
- Presented at the NAVER CLOVA AI Tech Talk (Nov 2021)