ASR and Text classification Pipeline

Text classification with Speech Recognition in unified pipeline

• We used Wav2Vec for speech recognition. If you want to know how to finetune wav2vec, please see here(for korean, see here)

• We used Electra(especially KoElectra as we worked on Korean dataset) for text classification. If you want to know how to finetune electra, please see here: for Korean

Installation

pip install -r requirements.txt

Inference

Assume you have both wav2vec and electra.

python main.py [-h] [--wav_dir WAV_DIR] [--stt_output_path STT_OUTPUT_PATH] [--output_path OUTPUT_PATH] [--wav2vec_checkpoint WAV2VEC_CHECKPOINT] [--electra_checkpoint ELECTRA_CHECKPOINT]

optional arguments:
  -h, --help            show this help message and exit
  --wav_dir WAV_DIR
  --stt_output_path STT_OUTPUT_PATH
  --output_path OUTPUT_PATH
  --wav2vec_checkpoint WAV2VEC_CHECKPOINT
  --electra_checkpoint ELECTRA_CHECKPOIN

The STT results will be in --stt_output_path. Final predicted output will be in --output_path.