RESULTS

Using Conformer encoder with Hubert pre-encoder, SpecAugment, speed perturbation and 3 context utterances

ASR config: conf/tuning/train_asr_conformer_hubert.yaml
Pretrained model:
- Zenodo: https://zenodo.org/record/5817199#.YdQ9_YTMKkA
- Hugging Face Hub: https://huggingface.co/espnet/akreal_swbd_da_hubert_conformer

Dataset	Snt	Dialogue Act Classification (%)
decode_asr_asr_model_valid.loss.ave/test	2379	66.3
decode_asr_asr_model_valid.loss.ave/valid	8117	69.5

Dataset	Snt	Dialogue Act Classification (%)
decode_asr_asr_model_valid.loss.ave/test	2379	52.9
decode_asr_asr_model_valid.loss.ave/valid	8117	56.1

Dataset	Snt	Dialogue Act Classification (%)
decode_asr_asr_model_valid.acc.best/test	2379	51.9
decode_asr_asr_model_valid.acc.best/valid	8117	56.8

Dataset	Snt	Dialogue Act Classification (%)
decode_asr_asr_model_valid.acc.best/test	2379	35.9
decode_asr_asr_model_valid.acc.best/valid	8117	39.4